Fine-tuning-as-a-service, while commercially successful for Large Language Model (LLM) providers, exposes models to harmful finetuning attacks.As a widely explored defense paradigm against such attacks, unlearning attempts to remove malicious knowledge from LLMs, thereby essentially preventing them from being used to perform malicious tasks.However, we highlight a critical flaw: the inherent general adaptability of LLMs allows them to easily bypass selective unlearning by rapidly relearning or repurposing their general capabilities for harmful tasks.To address this fundamental limitation, we propose a paradigm shift: instead of selective removal, we advocate for inducing model collapse, effectively forcing the model to "unlearn everything", specifically in response to updates characteristic of malicious adaptation.This collapse directly neutralizes the very general capabilities that attackers exploit, tackling the core issue unaddressed by selective unlearning.We introduce the Collapse Trap (CTRAP) as a practical mechanism to implement this concept conditionally.Embedded during alignment, CTRAP pre-configures the model's reaction to subsequent fine-tuning dynamics.If updates during fine-tuning constitute a persistent attempt to reverse safety alignment, the pre-configured trap triggers a progressive degradation of the model's core language modeling abilities, ultimately rendering it inert and useless for the attacker.Crucially, this collapse mechanism remains dormant during benign fine-tuning, ensuring the model's utility and general capabilities are preserved.1

Variance reduction has been shown to improve the performance of Stochastic Gradient Descent (SGD) in centralized machine learning. However, when it is extended to federated learning systems, many issues may arise, including (i) mega-batch size settings; (ii) additional noise introduced by the gradient difference between the current iteration and the snapshot point; and (iii) gradient (statistical) heterogeneity. In this paper, we propose a lightweight algorithm termed federated adaptive batch size time evolving variance reduction (FedATEVR) to tackle these issues, consisting of an adaptive batch size setting scheme and a time-evolving variance reduction gradient estimator. In particular, we use the historical gradient information to set an appropriate mega-batch size for each client, which can steadily accelerate the local SGD process and reduce the computation cost. The historical information involves both global and local gradient, which mitigates unstable varying in mega-batch size introduced by gradient heterogeneity among the clients. For each client, the gradient difference between the current iteration and the snapshot point is used to tune the time-evolving weight of the variance reduction term in the gradient estimator. This can avoid meaningless variance reduction caused by the out-of-date snapshot point gradient. We theoretically prove that our algorithm can achieve a linear speedup of of $\mathcal {O}(\frac{1}{\sqrt{SKT}})$O(1SKT) for non-convex objective functions under partial client participation. Extensive experiments demonstrate that our proposed method can achieve higher test accuracy than the baselines and decrease communication rounds greatly.

Recentworks have shown the potential of diffusion models in computer vision and natural language processing. Apart from the classical supervised learning fields, diffusion models have also shown strong competitiveness in reinforcement learning (RL) by formulating decision-making as sequential generation. However, incorporating temporal information of sequential data and utilizing it to guide diffusion models to perform better generation is still an open challenge. In this paper, we take one step forward to investigate controllable generation with temporal conditions that are refined from temporal information. We observe the importance of temporal conditions in sequential generation in sufficient scenarios and provide a comprehensive discussion and comparison of different temporal conditions. Based on the observations, we propose an effective temporally-conditional diffusion model coined Temporally-Composable Diffuser (TCD), which extracts temporal information from interaction sequences and explicitly guides generation with temporal conditions. Specifically, we separate the sequences into three parts according to time expansion and identify historical, immediate, and prospective conditions accordingly. Each condition preserves non-overlapping temporal information of sequences, enabling more controllable generation when we jointly use them to guide the diffuser. Finally, we conduct extensive experiments and analysis to reveal the favorable applicability of TCD in offline RL tasks, where our method reaches or matches the best performance compared with prior SOTA baselines.

Data-Free Meta-Learning (DFML) aims to enable efficient learning of unseen few-shot tasks, by meta-learning from multiple pre-trained models without accessing their original training data. While existing DFML methods typically generate synthetic data from these models to perform meta-learning, a comprehensive analysis of DFML's robustness-particularly its failure modes and vulnerability to potential attacks-remains notably absent. Such an analysis is crucial as algorithms often operate in complex and uncertain real-world environments. This paper fills this significant gap by systematically investigating the robustness of DFML, identifying two critical but previously overlooked vulnerabilities: Task-Distribution Shift (TDS) and Task-Distribution Corruption (TDC). TDS refers to the sequential shifts in the evolving task distribution, leading to the catastrophic forgetting of previously learned meta-knowledge. TDC exposes a security flaw of DFML, revealing its susceptibility to attacks when the pre-trained model pool includes untrustworthy models that deceptively claim to be beneficial but are actually harmful. To mitigate these vulnerabilities, we propose a trustworthy DFML framework comprising three components: synthetic task reconstruction, meta-learning with task memory interpolation, and automatic model selection. Specifically, utilizing model inversion techniques, we reconstruct synthetic tasks from multiple pre-trained models to perform meta-learning. To prevent forgetting, we introduce a strategy to replay interpolated historical tasks to efficiently recall previous meta-knowledge. Furthermore, our framework seamlessly incorporates an automatic model selection mechanism to automatically filter out untrustworthy models during the meta-learning process. Extensive experiments across various datasets with two types of untrustworthy models confirm the superiority of our method in significantly enhancing the robustness of DFML.

Decentralized Federated Learning has emerged as an alternative to centralized architectures due to its faster training, privacy preservation, and reduced communication overhead. In decentralized communication, the server aggregation phase in Centralized Federated Learning shifts to the client side, which means that clients connect with each other in a peer-to-peer manner. However, compared to the centralized mode, data heterogeneity in Decentralized Federated Learning will cause larger variances between aggregated models, which leads to slow convergence in training and poor generalization performance in tests. To address these issues, we introduce Catalyst Acceleration and propose an acceleration Decentralized Federated Learning algorithm called DFedCata. It consists of two main components: the Moreau envelope function, which primarily addresses parameter inconsistencies among clients caused by data heterogeneity, and Nesterov's extrapolation step, which accelerates the aggregation phase. Theoretically, we prove the optimization error bound and generalization error bound of the algorithm, providing a further understanding of the nature of the algorithm and the theoretical perspectives on the hyperparameter choice. Empirically, we demonstrate the advantages of the proposed algorithm in both convergence speed, computational cost, and generalization performance on CIFAR10/100 and Tiny-ImageNet with various non-iid data distributions. Moreover, extensive experiments are conducted to validate the theoretical properties of DFedCata, showing strong consistency between theory and empirical observations.

Multi-view data encompasses various data types, including multi-feature, multi-sequence, and multi-modal data. Multi-view multi-label classification aims to leverage the rich semantic information contained in multiple views to achieve enhanced multi-label classification performance. In practical applications, the absence of views and labels poses a significant challenge to multi-view multi-label classification tasks. Premised on the assumption that shared semantic information across multiple views is sufficient to support the downstream task, we propose CTRL, a novel incomplete multi-view multi-label classification framework to address the multi-view learning challenge on the data with partially missing views and missing labels in this paper. The core mechanism of CTRL lies in learning a high-purity, low-redundancy condensed representation that adequately captures the essential information of the original data. Specifically, we design a new objective loss to enhance the semantic information of shared cross-view within the joint representation learning process while simultaneously suppressing intra-view redundant information that is irrelevant to the downstream task. This enables CTRL to extract task-relevant representations even when views are incomplete. Furthermore, we employ the Beta Evidential Neural Network to model the label distribution. This network is then integrated with Dempster-Shafer theory, enabling our model to perform label-level classification uncertainty estimation. This also allows us to use the estimated uncertainty and belief mass to create high-reliability pseudo-labels, resulting in further gains in model performance. Experimental results on multiple benchmark datasets demonstrate the superior performance of our proposed model in terms of accuracy, robustness, and reliability.

Multi-task learning (MTL) leverages a shared model to accomplish multiple tasks and facilitate knowledge transfer. Recent research on task arithmetic-based MTL demonstrates that merging the parameters of independently fine-tuned models can effectively achieve MTL. However, existing merging methods primarily seek a static optimal solution within the original model parameter space, which often results in performance degradation due to the inherent diversity among tasks and potential interferences. To address this challenge, in this paper, we propose a Weight-Ensembling Mixture of Experts (WEMoE) method for multi-task model merging. Specifically, we first identify critical (or sensitive) modules by analyzing parameter variations in core modules of Transformer-based models before and after fine-tuning. Then, our WEMoE statically merges non-critical modules while transforming critical modules into a mixture-of-experts (MoE) structure. During inference, expert modules in the MoE are dynamically merged based on input samples, enabling a more flexible and adaptive merging approach. Building on WEMoE, we further introduce an efficient-and-effective WEMoE (E-WEMoE) method, whose core mechanism involves eliminating non-essential elements in the critical modules of WEMoE and implementing shared routing across multiple MoE modules, thereby significantly reducing both the trainable parameters, the overall parameter count, and computational overhead of the merged model by WEMoE. Experimental results across various architectures and tasks demonstrate that both WEMoE and E-WEMoE outperform state-of-the-art (SOTA) model merging methods in terms of MTL performance, generalization, and robustness.

As two mainstream frameworks in federated learning (FL), both centralized and decentralized approaches have shown great application value in practical scenarios. However, existing studies do not provide sufficient evidence and clear guidance for analysis of which performs better in the FL community. Although decentralized methods have been proven to approach the comparable convergence of centralized with less communication, their test performance always falls short of expectations in empirical studies. To comprehensively and fairly compare their efficiency gaps in FL, in this paper, we explore their stability and generalization efficiency. Specifically, we prove that on the general smooth non-convex objectives, 1) centralized FL (CFL) always generalizes better than decentralized FL (DFL); 2) CFL achieves the best performance via adopting partial participation instead of full participation; and, 3) there is a necessary requirement for the topology in DFL to avoid performance collapse as the training scale increases. We also conduct extensive experiments on several common setups in FL to validate that our theoretical analysis is consistent with experimental phenomena and contextually valid in several general and practical scenarios.

Deep model training on extensive datasets is increasingly becoming cost-prohibitive, prompting the widespread adoption of deep model fusion techniques to leverage knowledge from pre-existing models. From simple weight averaging to more sophisticated methods like AdaMerging, model fusion effectively improves model performance and accelerates the development of new models. However, potential interference between parameters of individual models and the lack of interpretability in the fusion progress remain significant challenges. Existing methods often try to resolve the parameter interference issue by evaluating attributes of parameters, such as their magnitude or sign, or by parameter pruning. In this study, we begin by examining the fine-tuning of linear layers through the lens of subspace analysis and explicitly define parameter interference as an optimization problem to shed light on this subject. Subsequently, we introduce an innovative approach to model fusion called zero-shot Sparse MIxture of Low-rank Experts (SMILE) construction, which allows for the upscaling of source models into an MoE model without extra data or further training. Our approach relies on the observation that fine-tuning mostly keeps the important parts from the pre-training, but it uses less significant or unused areas to adapt to new tasks. Additionally, the issue of parameter interference, which is intrinsically challenging in the original parameter space, can be managed by expanding the dimensions. We conduct extensive experiments across diverse scenarios, such as image classification and text generation tasks, using full fine-tuning and LoRA fine-tuning, and we apply our method to large language models (CLIP models, Flan-T5 models, and Mistral-7B models), highlighting the adaptability and scalability of SMILE. For full fine-tuned models, about 50% additional parameters can achieve around 98-99% of the performance of eight individual fine-tuned ViT models, while for LoRA fine-tuned Flan-T5 models, maintaining 99% performance with only 2% extra parameters. Code is available athttps://github.com/tanganke/fusion_bench.

Continual learning (CL) focuses on learning non-stationary data distribution without forgetting previous knowledge. The most widely used memory-replay approaches are often prone to memory overfitting due to the limited memory diversity and hardness. Existing work mitigating memory overfitting either lacks data diversity or hardness or is hard to train. To address the above limitations and release the memory buffer potential, we view the memory buffer transformation from a new dynamic system perspective and propose a continuous and reversible memory transformation method. We introduce an adversarial optimization objective that jointly learns the CL model and memory transformer. Specifically, we present a deterministic continuous memory transformer (DCMT) to generate diverse memory data. Furthermore, we inject uncertainty into the transformation function and develop a stochastic continuous memory transformer (SCMT), which substantially enhances the diversity of the transformed memory buffer. The presented neural transformation approaches have significant advantages over existing ones: (1) they significantly increase the memory buffer diversity and hardness to overfit; (2) they are memory efficient without needing to make a replica of the memory data. Extensive experiments show a significant improvement with our approach compared to strong baselines.

As a prominent research topic, multi-view multi-label classification (MvMlC) aims to assign multiple labels to samples by integrating information from various perspectives. However, in real-world scenarios, MvMlC frequently faces the learning challenge of data with missing views and labels, typically resulting from sensor malfunctions, or the costly and time-consuming process of manual annotation. In addition, learning robust representations that are both consistent across views and specific to individual views remains a challenge. To address these issues, we propose a novel double incomplete multi-view multi-label classification framework based on Disentangling Consistent and Specific Information (DCSI). Specifically, we employ a dual-channel encoder with identical architecture but distinct objectives to extract cross-view consistent information and view-specific unique information from all views, respectively. Meanwhile, a view discriminator is constructed to decouple these two types of information, facilitating the extraction of pure consistent and specific information. Moreover, we meticulously design fusion strategies tailored to each representation type. Regarding consistent representations, we propose a dynamic-confidence-aware fusion mechanism that assesses the reliability of each view's representations in relation to the classification task, enabling the model to prioritize information from trustworthy representations. For specific representations, in light of their complementary rather than redundant property, we suggest treating such representations from each view equally to ensure fairness. Through experimental validation on five datasets, the results demonstrate that our method outperforms existing state-of-the-art methods.

Adversarial training (AT) has been shown to effectively enhance a network's resilience against adversarial attack. However, conventional AT, which relies on a fixed pre-specified perturbation budget, suffers from several limitations when training robust models. First, enforcing the same perturbation budget across networks with different capacities leads to varying levels of robustness disparity between natural and robust accuracies, which deviates from the desired outcome of a robust network. Second, because the perturbation budget is fixed throughout training, the attack strength fails to scale adaptively with the evolving robustness of the model. This mismatch often results in robust overfitting and further degradation of adversarial robustness. To address these limitations, we propose a novel technique called Strength-Adaptive Adversarial Training (SAAT). In SAAT, the adversary incorporates an adversarial-loss constraint to guide the generation of adversarial training data. This constraint allows the perturbation budget to adapt dynamically based on the current training state, which effectively mitigates robust overfitting. Moreover, by explicitly regulating the attack strength through the adversarial loss, SAAT enables precise control over the robustness disparity between natural accuracy and adversarial robustness. Extensive experiments demonstrate that SAAT substantially improves adversarial robustness over standard AT.

Few-step diffusion models enable efficient high-resolution image synthesis but struggle to align with specific downstream objectives due to limitations of existing reinforcement learning (RL) methods in low-step regimes with limited state spaces and suboptimal sample quality. To address this, we propose Stepwise Diffusion Policy Optimization (SDPO), a novel RL framework tailored for few-step diffusion models. SDPO introduces a dual-state trajectory sampling mechanism, tracking both noisy and predicted clean states at each step to provide dense reward feedback and enable low-variance, mixed-step optimization. For further efficiency, we develop a latent similarity-based dense reward prediction strategy to minimize costly dense reward queries. Leveraging these dense rewards, SDPO optimizes a dense reward difference learning objective that enables more frequent and granular policy updates. Additional refinements, including stepwise advantage estimates, temporal importance weighting, and step-shuffled gradient updates, further enhance long-term dependency, low-step priority, and gradient stability. Our experiments demonstrate that SDPO consistently delivers superior reward-aligned results across diverse few-step settings and tasks.

Minimax optimization is gaining increasing attention in modern machine learning applications. Driven by large-scale models and massive volumes of data collected from edge devices, as well as the concern to preserve client privacy, distributed minimax optimization algorithms become popular, such as Local Stochastic Gradient Descent Ascent (Local-SGDA), and Local Decentralized SGDA (Local-DSGDA). While most existing research on distributed minimax algorithms focuses on convergence rates and communication efficiency, their generalization performance remains largely unexplored, whereas generalization ability is a pivotal indicator for evaluating the holistic performance of a model when fed with unknown data. In this paper, we propose the stability-based generalization analytical framework for Distributed-SGDA, which unifies two popular distributed minimax algorithms including Local-SGDA and Local-DSGDA, and conduct a comprehensive analysis of stability error, generalization gap, and population risk across different metrics under various settings, e.g., (S)C-(S)C, PL-SC, and NC-NC cases. Our theoretical results reveal the trade-off between the generalization gap and optimization error and suggest hyperparameters choice to obtain the optimal population risk. Numerical experiments for Local-SGDA and Local-DSGDA validate the theoretical results.

After training on a specific source domain, models can leverage domain generalization (DG) techniques to achieve superior and broader performance on new, unseen target domains. Existing DG often utilizes contrastive learning to learn domain-invariant features. The goal of contrastive learning is to learn effective representations of data, causing samples from the same category to cluster together in feature space, while samples from different categories are dispersed. Traditional contrastive learning is limited to a finite set of contrastive pairs for DG. To handle this problem, we consider sampling from an infinite number of contrastive pairs using a mixture of von Mises-Fisher (vMF) distributions on the unit hypersphere. We propose a novel method called Probability-guided Contrastive Learning (PgCL), which selects contrastive pairs based on estimated data distributions of samples from each category in feature space. Additionally, we derive the exact analytical formula for the expected contrastive loss. We conduct an empirical investigation of the error bounds of PgCL and demonstrate its performance by comparing it with several leading methods across a range of DG datasets.

Accurate medical image segmentation plays a vital role in clinical diagnostics by facilitating the precise delineation of anatomical structures and pathological regions. However, the performance of existing segmentation methods is often constrained by the scarcity of high-quality annotated datasets, as manual labeling is both labor-intensive and reliant on domain-specific expertise. To address this limitation without requiring additional annotations, we propose a novel multimodal segmentation framework that leverages medical text annotations as an auxiliary modality to complement visual information. In particular, our approach introduces a learnable encoding strategy for joint distribution modeling of image and text, which enables discriminative fusion and effectively suppresses cross-modal redundancy. Moreover, we innovatively design a frequency-domain prompt encoder based on the discrete wavelet transform (DWT) to capture multi-frequency features, thereby significantly enhancing the model's ability to delineate fine-grained boundaries. Overall, our framework integrates cross-attention for effective cross-modal interaction, employs joint distribution modeling to enable discriminative and redundancy-reduced multimodal fusion, and incorporates auxiliary supervision to strengthen the learning of task-relevant features. Extensive experiments on nine public datasets across three clinical tasks-including cell, lung infection, and polyp segmentation-demonstrate that our method achieves competitive segmentation performance while maintaining favorable computational efficiency. Comprehensive ablation studies and feature distribution visualizations further validate the effectiveness and robustness of our proposed components. The code will be made publicly available at https://github.com/chenpeng052/MDFP.

In numerous artificial intelligence applications, the collaborative efforts of multiple intelligent agents are imperative for the successful attainment of target objectives. To enhance coordination among these agents, a distributed communication framework is often employed, wherein each agent must be capable of encoding information received from the environment and determining how to share it with other agents as required by the task at hand. However, indiscriminate information sharing among all agents can be resource-intensive, and the adoption of manually pre-defined communication architectures imposes constraints on inter-agent communication, thus limiting the potential for effective collaboration. Moreover, the communication framework often remains static during inference, which may result in sustained high resource consumption, as in most cases, only key decisions necessitate information sharing among agents. In this study, we propose a novel approach where the communication structure between agents is represented as a learnable graph.We frame this challenge as the task of identifying the optimal communication graph while allowing the architecture parameters to be updated through regular optimization, which requires a bi-level optimization process. By applying continuous relaxation to the graph structure and integrating attention mechanisms, our method, CommFormer, effectively optimizes the communication graph and simultaneously refines the architectural parameters via gradient descent in an end-to-end manner. Additionally, we introduce a temporal gating mechanism for each agent, enabling dynamic decisions on whether to receive shared information at a given time, based on current observations, thus improving decisionmaking efficiency. Comprehensive experiments conducted across a range of cooperative tasks demonstrate the robustness of our model. Our approach enables agents to develop more coordinated and sophisticated strategies, maintaining effectiveness even with varying agent counts.

Recently, the efficient deployment and acceleration of transformer-based pre-trained models (TPMs) on resource-constrained edge devices for multimedia services have gained significant interest. Although early exiting is a feasible solution, it may lead to extra computational cost and substantial performance degradation compared to the original models. To tackle these issues, we propose a framework termed EEformer, which incorporates global-local heads (GLHs) into intermediate layers to construct the early exiting dynamic neural network (EDNN). The GLH can efficiently extract global and local information from hidden states produced by the backbone layer, thereby achieving a better performance-efficiency trade-off for the EDNN. Moreover, we propose a novel progressive fine-tuning strategy to steadily improve the efficiency of the EDNN while maintaining its performance comparable to the original mode through three fine-tuning stages. We conduct extensive experiments on image classification and natural language processing tasks, demonstrating the superiority of the proposed framework. In particular, the proposed framework achieves 1.87× speed-up while maintaining 99.0% performance on the CIFAR-100 dataset, and 3.05× speed-up while maintaining 98.5% performance on the SST-2 dataset.

Deep model fusion/merging is an emerging technique that integrates parameters or predictions from multiple deep learning (DL) models into a unified framework. It combines the abilities of different models to compensate for the biases and errors of an individual model, improving overall performance. However, deep model fusion, especially on large-scale DL models such as large language models (LLMs) and foundation models, faces several challenges, including high computational cost and interference between different heterogeneous models. In order to understand it better, we present a comprehensive survey to summarize the recent progress. We categorize existing model fusion methods as fourfold: 1) weight average (WA) averages the parameters of multiple models to obtain results closer to the optimal solution; 2) considering that direct averaging of models often yields suboptimal results, "mode connectivity" connects networks via paths of nonincreasing loss in weight spaces before the fusion. Along these paths, initial models are transformed into forms with consistent functions and better fusion effects; 3) similarly, for models with poor direct fusion results, "alignment" matches the corresponding units and merges these models, thus fully exploiting the corresponding relationships between the models; and 4) in addition to the above-mentioned methods of parameter fusion, "ensemble learning" fuses the outputs of multiple models in the inference stage to improve the accuracy and robustness of networks. In addition, we analyze the challenges of deep model fusion and illuminate the possible research directions in the future.

Compared with general Federated Learning (FL), Decentralized FL (DFL) has diminished central communication burdens and lower risks of disruption. To be compatible with real-world scenarios, some existing works produce several local personalized models rather than a universal model for all edge devices. However, they still suffer from inferior performance due to the full model aggregation in heterogeneous datasets. Therefore, we propose a DFL framework DFedMDC through partial model personalization, which can adapt to resource-heterogeneous environments (e.g., the Internet of Things (IoT)). Specifically, it personalizes the “right” components of local models and trains the shared modeluiand personal modelvialternately in each clienti. To further accelerate the convergence process, we propose DFedSMDC with properly directed noise perturbation into the gradient update. Theoretically, we provide convergence analysis of both algorithms in the general non-convex setting. It can shed light on how vital factors affect the convergence rate, such as these alternate updates of partial modelsui,vi, data heterogeneity δ2as well as various communication topologies (characterized by the spectral gap 1 − λ). Empirically, we confirm the state-of-the-art (SOTA) superiority of the proposed methods on several real-world datasets with various data distributions, relative to both SOTA personalized FL (PFL) and DFL baselines.

To address the communication challenges associated with Federated Learning (FL), Decentralized Federated Learning (DFL) eliminates the central server and trains the model with decentralized method, enabling each client to only communicate with its neighbors. However, per our analysis, model trained with DFL experiences performance degradation because of data-heterogeneity and time-varying topologies. To address these issues, we propose a Dynamic K-step Gradient Tracking (DKGT) method to enhance the performance of DFL over time varying networks. Specifically, DKGT employs K-step local updates and gradient tracking to reduce the communication cost and the variance from heterogeneous data distribution, and we use dynamic gradient tracking parameter to correct gradient over time varying graph. Theoretically, we derive a universal convergence rate for smooth and non-convex problem at the rate of$\mathcal{O}\left(\frac{\left(f(\textbf{x}_0)-f(\textbf{x}^*)\right)}{\sqrt{T}(L\sqrt{KN})^{-1}-\tau(pKL\sqrt{TKN})^{-1}}+\frac{\sigma^2}{KTN\tau(pK-\tau)}\right)$, that τ and p respectively represent the time window length and the connectivity of time-varying networks. Experimentally, we illustrate the robustness and effectiveness of this heterogeneity correction on extensive non-convex neural network training tasks over different topologies and dynamic network settings.

Multimodal large language models (MLLMs) have experienced significant advancements recently, but still struggle to recognize and interpret intricate details in high-resolution (HR) images effectively. While state-of-the-art (SOTA) MLLMs claim to process images at 4K resolution, existing MLLM benchmarks only support up to 2K, leaving the capabilities of SOTA models on true HR images largely untested. Furthermore, existing methods for enhancing HR image perception in MLLMs rely on computationally expensive visual instruction tuning. To address these limitations, we introduce HR-Bench, the first deliberately designed benchmark to rigorously evaluate MLLM performance on 4K & 8K images. Through extensive experiments, we demonstrate that while downsampling HR images leads to vision information loss, leveraging complementary modalities, e.g., text, can effectively compensate for this loss. Building upon this insight, we propose Divide, Conquer and Combine, a novel training-free framework for enhancing MLLM perception of HR images. Our method follows a three-staged approach: 1) Divide: recursively partitioning the HR image into patches and merging similar patches to minimize computational overhead, 2) Conquer: leveraging the MLLM to generate accurate textual descriptions for each image patch, and 3) Combine: utilizing the generated text descriptions to enhance the MLLM's understanding of the overall HR image. Extensive experiments show that: 1) the SOTA MLLM achieves 63% accuracy, which is markedly lower than the 87% accuracy achieved by humans on HR-Bench; 2) our method brings consistent and significant improvements (a relative increase of +6% on HR-Bench and +8% on general multimodal benchmarks).

Large Language Models (LLMs) such as ChatGPT demonstrate strong few-shot adaptability without requiring fine-tuning, positioning them ideal for data-limited and real-time applications. However, this adaptability has not yet been replicated in current Visual Foundation Models (VFMs), which require explicit fine-tuning with sufficient tuning data. Besides, the pretraining-finetuning paradigm has led to the surge of numerous task-specific modular components, such as Low-Rank Adaptation (LoRA). For the first time, we explore the potential of reusing diverse pre-tuned LoRAs without accessing their original training data, to achieve tuning-free few-shot adaptation in VFMs. Our framework, LoRA Recycle, distills a meta-LoRA from diverse pre-tuned LoRAs with a meta-learning objective, using synthetic data inversely generated from pre-tuned Lo-RAs themselves. The VFM, once equipped with the meta-LoRA, is empowered to solve new few-shot tasks in a single forward pass, akin to the in-context learning of LLMs. Additionally, we incorporate a double-efficient mechanism, accelerating the data-generation and meta-training process while maintaining or even improving performance. Extensive experiments across various few-shot classification benchmarks across both in- and cross-domain scenarios demonstrate the superiority of our framework. Code is available at https://github.com/Egg-Hu/LoRA-Recycle.

Weight decay is a widely used technique in training machine learning models, known to empirically enhance the generalization of Stochastic Gradient Descent (SGD). While intuitively weight decay allows SGD to train a regularized model rather than the original one, there is limited theoretical understanding of why SGD with weight decay (SGDW) yields results consistent with the unregularized model, or how weight decay improves generalization. This paper establishes a convergence theory for SGDW in the context of the unregularized model, under weaker assumptions than previous analyses of weight decay. Our theory demonstrates that weight decay does not accelerate the convergence of SGD. For generalization, we provide the first theoretical proof of weight decay’s benefit in nonconvex optimization. Additionally, we extend our results to sign-based stochastic gradient algorithms, such as SignSGD. Numerical experiments on classical benchmarks validate our theoretical findings.

The in-context learning (ICL) capability of large language models (LLMs) enables them to perform challenging tasks using provided demonstrations. However, ICL is highly sensitive to the ordering of demonstrations, leading to instability in predictions. This paper shows that this vulnerability can be exploited to design a natural attack—difficult for model providers to detect—that achieves nearly 80% success rate on LLaMA-3 by simply permuting the demonstrations. Existing mitigation methods primarily rely on post-processing and fail to enhance the model's inherent robustness to input permutations, raising concerns about safety and reliability of LLMs. To address this issue, we propose Permutation-resilient learning (PEARL), a novel framework based on distributionally robust optimization (DRO), which optimizes model performance against the worst-case input permutation. Specifically, PEARL consists of a permutation-proposal network (P-Net) and the LLM. The P-Net generates the most challenging permutations by treating it as an optimal transport problem, which is solved using an entropy-constrained Sinkhorn algorithm. Through minimax optimization, the P-Net and the LLM iteratively optimize against each other, progressively improving the LLM's robustness. Experiments on synthetic pre-training and real-world instruction tuning tasks demonstrate that PEARL effectively mitigates permutation attacks and enhances performance. Notably, despite being trained on fewer shots and shorter contexts, PEARL achieves performance gains of up to 40% when scaled to many-shot and long-context scenarios, highlighting its efficiency and generalization capabilities.

Selecting high-quality pre-training data for large language models (LLMs) is crucial for enhancing their overall performance under limited computation budget, improving both training and sample efficiency. Recent advancements in file selection primarily rely on using an existing or trained proxy model to assess the similarity of samples to a target domain, such as high quality sources BookCorpus and Wikipedia. However, upon revisiting these methods, the domain-similarity selection criteria demonstrates a diversity dilemma, i.e. dimensional collapse in the feature space, improving performance on the domain-related tasks but causing severe degradation on generic performance.To prevent collapse and enhance diversity, we propose a DiverSified File selection algorithm (DiSF), which selects the most decorrelated text files in the feature space. We approach this with a classical greedy algorithm to achieve more uniform eigenvalues in the feature covariance matrix of the selected texts, analyzing its approximation to the optimal solution under a formulation of $\gamma$-weakly submodular optimization problem. Empirically, we establish a benchmark and conduct extensive experiments on the TinyLlama architecture with models from 120M to 1.1B parameters. Evaluating across nine tasks from the Harness framework, DiSF demonstrates a significant improvement on overall performance. Specifically, DiSF saves 98.5\% of 590M training files in SlimPajama, outperforming the full-data pre-training within a 50B training budget, and achieving about 1.5x training efficiency and 5x data efficiency. Source code is available at: https://github.com/MediaBrain-SJTU/DiSF.git.

Despite great achievements in algorithm design for Personalized Federated Learning (PFL), research on the theoretical analysis of generalization is still in its early stages. Some theoretical results have investigated the generalization performance of personalized models under the problem setting and hypothesis in convex conditions, which can not reflect the real iteration performance during non-convex training. To further understand the real performance from a generalization perspective, we propose the first algorithm-dependent generalization analysis with uniform stability for the typical PFL method, Partial Model Personalization, on smooth and non-convex objectives. Specifically, we decompose the generalization errors into aggregation errors and fine-tuning errors, then creatively establish a generalization analysis framework corresponding to the gradient estimation process of the personalized training. This framework builds up the bridge among PFL, FL and Pure Local Training for personalized aims in heterogeneous scenarios, which clearly demonstrates the effectiveness of PFL from the generalization perspective. Moreover, we demonstrate the impact of trivial factors like learning steps, stepsizes and communication topologies and obtain the excess risk analysis with optimization errors for PFL. Promising experiments on CIFAR datasets also corroborate our theoretical insights. Our code can be seen in https://github.com/YingqiLiu1999/Understanding-the-Stability-based-Generalization-of-Personalized-Federated-Learning.

Model extraction aims to acquire a pre-trained black-box model concealed behind a black-box API. Existing defense strategies against model extraction primarily concentrate on preventing the unauthorized extraction of API functionality. However, two significant challenges still need to be solved: (i) Neural network architecture of the API constitutes a form of intellectual property that also requires protection; (ii) The current practice of allocating the same network architecture to both attack and benign queries results in substantial resource wastage. To address these challenges, we propose a novel \textit{Dynamic Neural Fortresses} (DNF) defense method, employing a dynamic Early-Exit neural network, deviating from the conventional fixed architecture. Firstly, we facilitate the random exit of attack queries from the network at earlier layers. This strategic exit point selection significantly reduces the computational cost for attack queries. Furthermore, the random exit of attack queries from earlier layers introduces increased uncertainty for attackers attempting to discern the exact architecture, thereby enhancing architectural protection. On the contrary, we aim to facilitate benign queries to exit at later layers, preserving model utility, as these layers typically yield meaningful information. Extensive experiments on defending against various model extraction scenarios and datasets demonstrate the effectiveness of DNF, achieving a notable 2$\times$ improvement in efficiency and an impressive reduction of up to 12\% in clone model accuracy compared to SOTA defense methods. Additionally, DNF provides strong protection against neural architecture theft, effectively safeguarding network architecture from being stolen.

Vision-language models such as CLIP have demonstrated strong zero-shot performance, but their considerable size and inefficient inference limit customizable deployment for users. While knowledge distillation is a solution, it still requires the original data, which is not always available due to copyrights and privacy concerns. For many users seeking open-vocabulary customization, Data-Free Knowledge Distillation (DFKD) emerges as a promising direction. Upon rethinking DFKD, we find that existing methods fail on CLIP due to their heavy reliance on BatchNorm layers, which are unexpectedly unusable in CLIP. Based on our findings, we adopt image-text matching to achieve DFKD for CLIP, enabling customization based on arbitrary class texts. This involves (i) inversing a surrogate dataset from CLIP based on text prompts; and (ii) distilling a student model from CLIP using the surrogate dataset. Specifically, we introduce style dictionary diversification to enhance the diversity of synthetic images. To prevent uncontrollable semantics introduced by diversification, we propose a class consistency maintaining strategy to ensure the consistency of synthetic images. Based on synthetic images with various styles, we further propose meta knowledge distillation to train the student model with good generalization ability. Moreover, we introduce a simple yet effective method to enable customization based on few example images. Comprehensive experiments showcase the superiority of our approach across twelve customized tasks, achieving a 9.33\% improvement compared to existing DFKD methods.

Model merging has gained significant attention as a cost-effective approach to integrate multiple single-task fine-tuned models into a unified one that can perform well on multiple tasks. However, existing model merging techniques primarily focus on resolving conflicts between task-specific models, they often overlook potential security threats, particularly the risk of backdoor attacks in the open-source model ecosystem. In this paper, we first investigate the vulnerabilities of existing model merging methods to backdoor attacks, identifying two critical challenges: backdoor succession and backdoor transfer. To address these issues, we propose a novel Defense-Aware Merging (DAM) approach that simultaneously mitigates task interference and backdoor vulnerabilities. Specifically, DAM employs a meta-learning-based optimization method with dual masks to identify a shared and safety-aware subspace for model merging. These masks are alternately optimized: the Task-Shared mask identifies common beneficial parameters across tasks, aiming to preserve task-specific knowledge while reducing interference, while the Backdoor-Detection mask isolates potentially harmful parameters to neutralize security threats. This dual-mask design allows us to carefully balance the preservation of useful knowledge and the removal of potential vulnerabilities. Compared to existing merging methods, DAM achieves a more favorable balance between performance and security, reducing the attack success rate by 2-10 percentage points while sacrificing only about 1\% in accuracy. Furthermore, DAM exhibits robust performance and broad applicability across various types of backdoor attacks and the number of compromised models involved in the merging process. Our codes and models can be accessed through https://github.com/Yangjinluan/DAM.

Coordinating multiple agents to collaboratively maximize submodular functions in unpredictable environments is a critical task with numerous applications in machine learning, robot planning and control. The existing approaches, such as the OSG algorithm, are often hindered by their poor approximation guarantees and the rigid requirement for a fully connected communication graph. To address these challenges, we firstly present a $\textbf{MA-OSMA}$ algorithm, which employs the multi-linear extension to transfer the discrete submodular maximization problem into a continuous optimization, thereby allowing us to reduce the strict dependence on a complete graph through consensus techniques. Moreover, $\textbf{MA-OSMA}$ leverages a novel surrogate gradient to avoid sub-optimal stationary points. To eliminate the computationally intensive projection operations in $\textbf{MA-OSMA}$, we also introduce a projection-free $\textbf{MA-OSEA}$ algorithm, which effectively utilizes the KL divergence by mixing a uniform distribution. Theoretically, we confirm that both algorithms achieve a regret bound of $\widetilde{O}(\sqrt{\frac{C_{T}T}{1-\beta}})$ against a  $(\frac{1-e^{-c}}{c})$-approximation to the best comparator in hindsight, where $C_{T}$ is the deviation of maximizer sequence, $\beta$ is the spectral gap of the network and $c$ is the joint curvature of submodular objectives. This result significantly improves the $(\frac{1}{1+c})$-approximation provided by the state-of-the-art OSG algorithm. Finally, we demonstrate the effectiveness of our proposed algorithms through simulation-based multi-target tracking.

Label differential privacy (DP) is a framework that protects the privacy of labels in training datasets, while the feature vectors are public. Existing approaches protect the privacy of labels by flipping them randomly, and then train a model to make the output approximate the privatized label. However, as the number of classes K increases, stronger randomization is needed, thus the performances of these methods become significantly worse. In this paper, we propose a vector approximation approach for learning with label local differential privacy, which is easy to implement and introduces little additional computational overhead. Instead of flipping each label into a single scalar, our method converts each label into a random vector with K components, whose expectations reflect class conditional probabilities. Intuitively, vector approximation retains more information than scalar labels. A brief theoretical analysis shows that the performance of our method only decays slightly with K. Finally, we conduct experiments on both synthesized and real datasets, which validate our theoretical analysis as well as the practical performance of our method.

Harmful fine-tuning (HFT), performed directly on open-source LLMs or through Fine-tuning-as-a-Service, breaks safety alignment and poses significant threats. Existing methods aim to mitigate HFT risks by learning robust representation on alignment data or making harmful data unlearnable, but they treat each data sample equally, leaving data vulnerability patterns understudied. In this work, we reveal that certain subsets of alignment data are consistently more prone to forgetting during HFT across different fine-tuning tasks and exhibit lower robustness compared to other subsets. Inspired by these findings, we propose Vulnerability-Aware Alignment (VAA), which calculates data vulnerability, partitions data into "vulnerable" and "invulnerable" groups, and encourages balanced learning using a group distributionally robust optimization (Group DRO) framework. Specifically, VAA learns an adversarial sampler that samples examples from the currently underperforming group and then applies group-dependent adversarial perturbations to the data during training, aiming to encourage a balanced learning process across groups. Experiments across four fine-tuning tasks demonstrate that VAA significantly reduces harmful scores while preserving downstream task performance, outperforming state-of-the-art baselines.

Identifying the most representative subset for a close-to-submodular objective while satisfying the predefined partition constraint is a fundamental task with numerous applications in machine learning. However, the existing distorted local-search methods are often hindered by their prohibitive query complexities and the rigid requirement for prior knowledge of difficult-to-obtain structural parameters. To overcome these limitations, we introduce a novel algorithm titled **Multinoulli-SCG**, which not only is parameter-free, but also can achieve the same approximation guarantees as the distorted local-search methods with significantly fewer function evaluations. The core of our **Multinoulli-SCG** algorithm is an innovative continuous-relaxation framework named Multinoulli Extension(***ME***), which can effectively convert the discrete subset selection problem subject to partition constraints into a solvable continuous maximization focused on learning the optimal multinoulli priors across the considered partition. In sharp contrast with the well-established multi-linear extension for submodular subset selection, a notable advantage of our proposed ***ME*** is its intrinsic capacity to provide a lossless rounding scheme for any set function. Finally, we validate the practical efficacy of our proposed algorithms by applying them to video summarization, bayesian A-optimal design and coverage maximization.

Pruning is a widely used technique for compressing large neural networks that eliminates weights that have minimal impact on the model's performance. Current pruning methods, exemplified by magnitude pruning, assign an importance score to each weight based on its magnitude and remove weights with scores below a certain threshold. Nonetheless, these methods often create a gap between the original dense and the pruned sparse model, potentially impairing performance. Especially when the sparsity ratio is high, the gap becomes more pronounced. To mitigate this issue, we introduce a method to bridge the gap left by pruning by utilizing a low-rank approximation of the difference between the dense and sparse matrices. Our method entails the iterative refinement of the sparse weight matrix augmented by a low-rank adjustment. This technique captures and retains the essential information often lost during pruning, thereby improving the performance of the pruned model. Furthermore, we offer a comprehensive theoretical analysis of our approach, emphasizing its convergence properties and establishing a solid basis for its efficacy. Experimental results on LLaMa models validate its effectiveness on large language models across various pruning techniques and sparsity levels. Our method shows significant improvements: at 50\% sparsity, it reduces perplexity by 53.9\% compared to conventional magnitude pruning on LLaMa-7B. Furthermore, to achieve a specific performance target, our approach enables an 8.6\% reduction in model parameters while maintaining a sparsity ratio of about 50\%.

Despite recent advancements in offline multi-task reinforcement learning (MTRL) have harnessed the powerful capabilities of the Transformer architecture, most approaches focus on a limited number of tasks, with scaling to extremely massive tasks remaining a formidable challenge. In this paper, we first revisit the key impact of task numbers on current MTRL method, and further reveal that naively expanding the parameters proves insufficient to counteract the performance degradation as the number of tasks escalates. Building upon these insights, we propose M3DT, a novel mixture-of-experts (MoE) framework that tackles task scalability by further unlocking the model’s parameter scalability. Specifically, we enhance both the architecture and the optimization of the agent, where we strengthen the Decision Transformer (DT) backbone with MoE to reduce task load on parameter subsets, and introduce a three-stage training mechanism to facilitate efficient training with optimal performance. Experimental results show that, by increasing the number of experts, M3DT not only consistently enhances its performance as model expansion on the fixed task numbers, but also exhibits remarkable task scalability, successfully extending to 160 tasks with superior performance.

Effectively scaling up deep reinforcement learning models has proven notoriously difficult due to network pathologies during training, motivating various targeted interventions such as periodic reset and architectural advances such as layer normalization. Instead of pursuing more complex modifications, we show that introducing static network sparsity alone can unlock further scaling potential beyond their dense counterparts with state-of-the-art architectures. This is achieved through simple one-shot random pruning, where a predetermined percentage of network weights are randomly removed once before training. Our analysis reveals that, in contrast to naively scaling up dense DRL networks, such sparse networks achieve both higher parameter efficiency for network expressivity and stronger resistance to optimization challenges like plasticity loss and gradient interference. We further extend our evaluation to visual and streaming RL scenarios, demonstrating the consistent benefits of network sparsity.

High-resolution (HR) image perception remains a key challenge in multimodal large language models (MLLMs). To drive progress beyond the limits of heuristic methods, this paper advances HR perception capabilities of MLLMs by harnessing cutting-edge long-context techniques such as retrieval-augmented generation (RAG). Towards this end, this paper presents the first study exploring the use of RAG to address HR perception challenges. Specifically, we propose Retrieval-Augmented Perception (RAP), a training-free framework that retrieves and fuses relevant image crops while preserving spatial context using the proposed Spatial-Awareness Layout. To accommodate different tasks, the proposed Retrieved-Exploration Search (RE-Search) dynamically selects the optimal number of crops based on model confidence and retrieval scores. Experimental results on HR benchmarks demonstrate the significant effectiveness of RAP, with LLaVA-v1.5-13B achieving a 43% improvement on $V^*$ Bench and 19% on HR-Bench. Code is available at https://github.com/DreamMr/RAP.

Training safe LLMs remains a critical challenge. The most widely used method, Refusal Training (RT), struggles to generalize against various Out-of-Distribution (OOD) jailbreaking attacks. Although various advanced methods have been proposed to address this issue, we instead question whether OOD attacks inherently surpass the capability of vanilla RT. Evaluations using Best-of-N (BoN) reveal significant safety improvements as N increases, indicating models possess adequate latent safety knowledge but RT fails to consistently elicit it under OOD scenarios. Further domain adaptation analysis reveals that direct RT causes reliance on superficial shortcuts, resulting in non-generalizable representation mappings. Inspired by our findings, we propose training model to perform safety reasoning for each query. Specifically, we synthesize reasoning supervision aligned with specified guidelines that reflect diverse perspectives on safety knowledge. This encourages model to engage in deeper reasoning, explicitly eliciting and utilizing latent safety knowledge for each query. Extensive experiments show that our method significantly improves model generalization against OOD attacks.

Semi-supervised learning (SSL) has made notable advancements in medical image segmentation (MIS), particularly in scenarios with limited labeled data and significantly enhancing data utilization efficiency. Previous methods primarily focus on complex training strategies to utilize unlabeled data but neglect the importance of graph structural information. Different from existing methods, we propose a graph-based clustering for semi-supervised medical image segmentation (GraphCL) by jointly modeling graph data structure in a unified deep model. The proposed GraphCL model enjoys several advantages. Firstly, to the best of our knowledge, this is the first work to model the data structure information for semi-supervised medical image segmentation (SSMIS). Secondly, to get the clustered features across different graphs, we integrate both pairwise affinities between local image features and raw features as inputs. Extensive experimental results on three standard benchmarks show that the proposed GraphCL algorithm outperforms state-of-the-art semi-supervised medical image segmentation methods.

Merging multiple expert models offers a promising approach for performing multi-task learning without accessing their original data. Existing methods attempt to alleviate task conflicts by sparsifying task vectors or promoting orthogonality among them. However, they overlook the fundamental target of model merging: the merged model performs as closely as possible to task-specific models on respective tasks. We find these methods inevitably discard task-specific information that, while causing conflicts, is crucial for performance. Based on our findings, we frame model merging as a constrained optimization problem ($\textit{i.e.}$, minimizing the gap between the merged model and individual models, subject to the constraint of retaining shared knowledge) and solve it via adaptive projective gradient descent. Specifically, we align the merged model with individual models by decomposing and reconstituting the loss function, alleviating conflicts through $\textit{data-free}$ optimization of task vectors. To retain shared knowledge, we optimize this objective by projecting gradients within a $\textit{shared subspace}$ spanning all tasks. Moreover, we view merging coefficients as adaptive learning rates and propose a task-aware, training-free strategy. Experiments show that our plug-and-play approach consistently outperforms previous methods, achieving state-of-the-art results across diverse architectures and tasks in both vision and NLP domains.

Nonparametric contextual bandit is an important model of sequential decision making problems. Under $\alpha$-Tsybakov margin condition, existing research has established a regret bound of $\tilde{O}\left(T^{1-\frac{\alpha+1}{d+2}}\right)$ for bounded supports. However, the optimal regret with unbounded contexts has not been analyzed. The challenge of solving contextual bandit problems with unbounded support is to achieve both exploration-exploitation tradeoff and bias-variance tradeoff simultaneously. In this paper, we solve the nonparametric contextual bandit problem with unbounded contexts. We propose two nearest neighbor methods combined with UCB exploration. The first method uses a fixed $k$. Our analysis shows that this method achieves minimax optimal regret under a weak margin condition and relatively light-tailed context distributions. The second method uses adaptive $k$. By a proper data-driven selection of $k$, this method achieves an expected regret of $\tilde{O}\left(T^{1-\frac{(\alpha+1)\beta}{\alpha+(d+2)\beta}}+T^{1-\beta}\right)$, in which $\beta$ is a parameter describing the tail strength. This bound matches the minimax lower bound up to logarithm factors, indicating that the second method is approximately optimal.

The Conditional Sequence Modeling (CSM) paradigm, benefiting from the transformer’s powerful distribution modeling capabilities, has demonstrated considerable promise in offline Reinforcement Learning (RL) tasks. Depending on the task’s nature, it is crucial to carefully balance the interplay between inherent local features and long-term dependencies in Markov decision trajectories to mitigate potential performance degradation and unnecessary computational overhead. In this paper, we propose Decision Mixer (DM), which addresses the conflict between features of different scales in the modeling process from the perspective of dynamic integration. Drawing inspiration from conditional computation, we design a plug-and-play dynamic token selection mechanism to ensure the model can effectively allocate attention to different features based on task characteristics. Additionally, we employ an auxiliary predictor to alleviate the short-sightedness issue in the autoregressive sampling process. DM achieves state-of-the-art performance on various standard RL benchmarks while requiring significantly fewer computational resources, offering a viable solution for building efficient and scalable RL foundation models. Code is available at here.

Large Language Models (LLMs) are discovered to suffer from accurately retrieving key information. To address this, we propose Mask-Enhanced Autoregressive Prediction (MEAP), a simple yet effective training paradigm that seamlessly integrates Masked Language Modeling (MLM) into Next-Token Prediction (NTP) to enhance the latter’s in-context retrieval capabilities. Specifically, MEAP first randomly masks a small fraction of input tokens and then directly performs the standard next-token prediction autoregressive using a decoder-only Transformer. MEAP eliminates the need for bidirectional attention or encoder-decoder architectures for MLM, incurring no additional computational overhead during pre-training or inference. Intensive experiments demonstrate that MEAP substantially outperforms NTP on key information retrieval and long-context reasoning tasks, while performing on par or better on commonsense reasoning tasks. The benefits of MEAP also extend to supervised fine-tuning, where it shows remarkable advantages in lost-in-the-middle scenarios, outperforming NTP by 11.77% percentage points. Our analysis indicates that MEAP’s effectiveness arises from its ability to promote more distinguishable attention scores by concentrating on a reduced set of non-masked tokens. This mechanism improves the model’s focus on task-relevant signals while mitigating the influence of peripheral context. These findings position MEAP as a promising training paradigm for large language models. Code has been submitted.

Personalized Federated Learning (PFL) meets each user’s personalized needs while still facing the high communication costs due to the large amount of data transmission and frequent communication. Decentralized PFL (DPFL) as an alternative discards the central server in PFL, which reduces the pressure of communication and the risk of server failure by using peer-to-peer communication.Nevertheless, DPFL still suffers from the significant communication pressure due to the transmission of a large number of model parameters, especially numerous nodes. To address the issues, we propose a novel personalized framework, DFedHP, in which each client utilizes a hypernetwork to generate the shared part of model parameters and train the personalized parameters separately. The number of parameters in a hypernetwork is much smaller than those in a typical local network, so hypernetwork aggregation reduces communication costs and the risk of privacy leakage. Furthermore, DFedHP can seamlessly integrate into existing DPFL algorithms as a plugin to boost their efficacy. At last, extensive experiments on various data heterogeneous environments demonstrate that DFedHP can reduce communication costs, accelerate convergence rate, and improve generalization performance compared with state-of-the-art (SOTA) baselines.

In current web environment, fake news spreads rapidly across online social networks, posing serious threats to society. Existing multimodal fake news detection methods can generally be classified into knowledge-based and semantic-based approaches. However, these methods are heavily rely on human expertise and feedback, lacking flexibility. To address this challenge, we propose a Dynamic Analysis and Adaptive Discriminator (DAAD) approach for fake news detection. For knowledge-based methods, we introduce the Monte Carlo Tree Search algorithm to leverage the self-reflective capabilities of large language models (LLMs) for prompt optimization, providing richer, domain-specific details and guidance to the LLMs, while enabling more flexible integration of LLM comment on news content. For semantic-based methods, we define four typical deceit patterns: emotional exaggeration, logical inconsistency, image manipulation, and semantic inconsistency, to reveal the mechanisms behind fake news creation. To detect these patterns, we carefully design four discriminators and expand them in depth and breadth, using the soft-routing mechanism to explore optimal detection models. Experimental results on three real-world datasets demonstrate the superiority of our approach.

Large generative models are increasingly trained on synthetic data from earlier generations, raising concerns about *model collapse*, a progressive performance decline consistently observed in empirical studies. However, theoretical understanding of recursive training dynamics and their failure modes remains limited. In this work, we theoretically show that recursive training inherently leads to exponential error growth unless mitigated by sufficient real data. Addressing the growing scarcity of real data, we introduce a self-verification mechanism enabling models to filter their outputs based on internal confidence scores without external validation. Through rigorous analysis, we derive finite-sample error bounds demonstrating that self-verification alone can prevent collapse, even in fully synthetic training regimes. Our theoretical framework extends to large language models (LLMs), characterizing the conditions under which recursive training can maintain stability without performance degradation.

Recent advances in multimodal semantic segmentation show that incorporating auxiliary inputs—such as depth or thermal images—can significantly improve performance over single-modality (RGB-only) approaches. However, most existing solutions rely on parallel backbone networks and complex fusion modules, greatly increasing model size and computational demands. Inspired by prompt tuning in large language models, we introduce \textbf{MixPrompt}: a prompting-based framework that integrates auxiliary modalities into a pretrained RGB segmentation model without modifying its architecture. MixPrompt uses a lightweight prompting module to extract and fuse information from auxiliary inputs into the main RGB backbone. This module is initialized using the early layers of a pretrained RGB feature extractor, ensuring a strong starting point. At each backbone layer, MixPrompt aligns RGB and auxiliary features in multiple low-rank subspaces, maximizing information use with minimal parameter overhead. An information mixing scheme enables cross-subspace interaction for further performance gains. During training, only the prompting module and segmentation head are updated, keeping the RGB backbone frozen for parameter efficiency. Experiments across NYU Depth V2, SUN-RGBD, MFNet, and DELIVER datasets show that MixPrompt achieves improvements of 4.3, 1.1, 0.4, and 1.1 mIoU, respectively, over two-branch baselines, while using nearly half the parameters. MixPrompt also outperforms recent prompting-based methods under similar compute budgets.

Weight decay is a standard regularization technique for training large language models (LLMs). While it is common to assign a uniform decay rate to every layer, this approach overlooks the structural diversity of LLMs and the varying spectral properties across modules. In this paper, we introduce AlphaDecay, a simple yet effective method that adaptively assigns different weight decay strengths to each module of an LLM. Our approach is guided by Heavy-Tailed Self-Regularization (HT-SR) theory, which analyzes the empirical spectral density (ESD) of weight correlation matrices to quantify “heavy-tailedness.” Modules exhibiting more pronounced heavy-tailed ESDs, reflecting stronger feature learning, are assigned weaker decay, while modules with lighter-tailed spectra receive stronger decay. Our method leverages tailored weight decay assignments to balance the module-wise differences in spectral properties, leading to improved performance. Extensive pre-training tasks with various model sizes from 60M to 1B demonstrate that AlphaDecay achieves better perplexity and generalization than conventional uniform decay and other adaptive decay baselines. The code is available at https://github.com/hed-ucas/AlphaDecay.

Pretraining a policy on offline data followed by fine-tuning through online interactions, known as Offline-to-Online Reinforcement Learning (O2O RL), has emerged as a promising paradigm for real-world RL deployment. However, both offline datasets and online interactions in practical environments are often noisy or even maliciously corrupted, severely degrading the performance of O2O RL. Existing works primarily focus on mitigating the conservatism of offline policies via online exploration, while the robustness of O2O RL under data corruption, including states, actions, rewards, and dynamics, is still unexplored. In this work, we observe that data corruption induces heavy-tailed behavior in the policy, thereby substantially degrading the efficiency of online exploration. To address this issue, we incorporate Inverse Probability Weighted (IPW) into the online exploration policy to alleviate heavy-tailedness, and propose a novel, simple yet effective method termed $\textbf{RPEX}$: $\textbf{R}$obust $\textbf{P}$olicy $\textbf{EX}$pansion. Extensive experimental results on D4RL datasets demonstrate that RPEX achieves SOTA O2O performance across a wide range of data corruption scenarios.

Continual offline reinforcement learning (CORL) has shown impressive ability in diffusion-based continual learning systems by modeling the joint distributions of trajectories. However, most research only focuses on limited continual task settings where the tasks have the same observation and action space, which deviates from the realistic demands of training agents in various environments. In view of this, we propose Vector-Quantized Continual Diffuser, named VQ-CD, to break the barrier of different spaces between various tasks. Specifically, our method contains two complementary sections, where the quantization spaces alignment provides a unified basis for the selective weights activation. In the quantized spaces alignment, we leverage vector quantization to align the different state and action spaces of various tasks, facilitating continual training in the same space. Then, we propose to leverage a unified diffusion model attached by the inverse dynamic model to master all tasks by selectively activating different weights according to the task-related sparse masks. Finally, we conduct extensive experiments on 15 continual learning (CL) tasks, including conventional CL task settings (identical state and action spaces) and general CL task settings (various state and action spaces). Compared with 17 baselines, our method reaches the SOTA performance.

Conditional decision generation with diffusion models has shown powerful competitiveness in reinforcement learning (RL). Recent studies reveal the relation between energy-function-guidance diffusion models and constrained RL problems. The main challenge lies in estimating the intermediate energy, which is intractable due to the log-expectation formulation during the generation process. To address this issue, we propose the Analytic Energy-guided Policy Optimization (AEPO). Specifically, we first provide a theoretical analysis and the closed-form solution of the intermediate guidance when the diffusion model obeys the conditional Gaussian transformation. Then, we analyze the posterior Gaussian distribution in the log-expectation formulation and obtain the target estimation of the log-expectation under mild assumptions. Finally, we train an intermediate energy neural network to approach the target estimation of log-expectation formulation. We apply our method in 30+ offline RL tasks to demonstrate the effectiveness of our method. Extensive experiments illustrate that our method surpasses numerous representative baselines in D4RL offline reinforcement learning benchmarks.

Harmful fine-tuning poses critical safety risks to fine-tuning-as-a-service for large language models. Existing defense strategies preemptively build robustness via attack simulation but suffer from fundamental limitations: (i) the infeasibility of extending attack simulations beyond bounded threat models due to the inherent difficulty of anticipating unknown attacks, and (ii) limited adaptability to varying attack settings, as simulation fails to capture their variability and complexity. To address these challenges, we propose Bayesian Data Scheduler (BDS), an adaptive tuning-stage defense strategy with no need for attack simulation. BDS formulates harmful fine-tuning defense as a Bayesian inference problem, learning the posterior distribution of each data point's safety attribute, conditioned on the fine-tuning and alignment datasets. The fine-tuning process is then constrained by weighting data with their safety attributes sampled from the posterior, thus mitigating the influence of harmful data. By leveraging the post hoc nature of Bayesian inference, the posterior is conditioned on the fine-tuning dataset, enabling BDS to tailor its defense to the specific dataset, thereby achieving adaptive defense. Furthermore, we introduce a neural scheduler based on amortized Bayesian learning, enabling efficient transfer to new data without retraining. Comprehensive results across diverse attack and defense settings demonstrate the state-of-the-art performance of our approach. Code is available at https://github.com/Egg-Hu/Bayesian-Data-Scheduler.

Recent advancements in reasoning capability of Multimodal Large Language Models (MLLMs) demonstrate its effectiveness in tackling complex visual tasks. However, existing MLLM-based Video Anomaly Detection (VAD) methods remain limited to shallow anomaly descriptions without deep reasoning. In this paper, we propose a new task named Video Anomaly Reasoning (VAR), which aims to enable deep analysis and understanding of anomalies in the video by requiring MLLMs to think explicitly before answering. To this end, we propose Vad-R1, an end-to-end MLLM-based framework for VAR. Specifically, we design a Perception-to-Cognition Chain-of-Thought (P2C-CoT) that simulates the human process of recognizing anomalies, guiding the MLLMs to reason about anomalies step-by-step. Based on the structured P2C-CoT, we construct Vad-Reasoning, a dedicated dataset for VAR. Furthermore, we propose an improved reinforcement learning algorithm AVA-GRPO, which explicitly incentivizes the anomaly reasoning capability of MLLMs through a self-verification mechanism with limited annotations. Experimental results demonstrate that Vad-R1 achieves superior performance, outperforming both open-source and proprietary models on VAD and VAR tasks.

Decentralized training removes the centralized server, making it a communication-efficient approach that can significantly improve training efficiency, but it often suffers from degraded performance compared to centralized training. Multi-Gossip Steps (MGS) serve as a simple yet effective bridge between decentralized and centralized training, significantly reducing experiment performance gaps. However, the theoretical reasons for its effectiveness and whether this gap can be fully eliminated by MGS remain open questions. In this paper, we derive upper bounds on the generalization error and excess error of MGS using stability analysis, systematically answering these two key questions. 1). Optimization Error Reduction: MGS reduces the optimization error bound at an exponential rate, thereby exponentially tightening the generalization error bound and enabling convergence to better solutions. 2). Gap to Centralization: Even as MGS approaches infinity, a non-negligible gap in generalization error remains compared to centralized mini-batch SGD ($\mathcal{O}(T^{\frac{c\beta}{c\beta +1}}/{n m})$ in centralized and $\mathcal{O}(T^{\frac{2c\beta}{2c\beta +2}}/{n m^{\frac{1}{2c\beta +2}}})$ in decentralized). Furthermore, we provide the first unified analysis of how factors like learning rate, data heterogeneity, node count, per-node sample size, and communication topology impact the generalization of MGS under non-convex settings without the bounded gradients assumption, filling a critical theoretical gap in decentralized training. Finally, promising experiments on CIFAR datasets support our theoretical findings.

Recently, long-thought reasoning models achieve strong performance on complex reasoning tasks, but often incur substantial inference overhead, making efficiency a critical concern. Our empirical analysis reveals that the benefit of using Long-CoT varies across problems: while some problems require elaborate reasoning, others show no improvement—or even degraded accuracy. This motivates adaptive reasoning strategies that tailor reasoning depth to the input. However, prior work primarily reduces redundancy within long reasoning paths, limiting exploration of more efficient strategies beyond the Long-CoT paradigm. To address this, we propose a novel two-stage framework for adaptive and efficient reasoning. First, we construct a hybrid reasoning model by merging long and short CoT models to enable diverse reasoning styles. Second, we apply bi-level preference training to guide the model to select suitable reasoning styles (group-level), and prefer concise and correct reasoning within each style group (instance-level). Experiments demonstrate that our method significantly reduces inference costs compared to other baseline approaches, while maintaining performance. Notably, on five mathematical datasets, the average length of reasoning is reduced by more than 50\%, highlighting the potential of adaptive strategies to optimize reasoning efficiency in large language models.

Deep model merging represents an emerging research direction that combines multiple fine-tuned models to harness their specialized capabilities across different tasks and domains. Current model merging techniques focus on merging all available models simultaneously, with weight interpolation-based methods being the predominant approach. However, these conventional approaches are not well-suited for scenarios where models become available sequentially, and they often suffer from high memory requirements and potential interference between tasks. In this study, we propose a training-free projection-based continual merging method that processes models sequentially through orthogonal projections of weight matrices and adaptive scaling mechanisms. Our method operates by projecting new parameter updates onto subspaces orthogonal to existing merged parameter updates while using an adaptive scaling mechanism to maintain stable parameter distances, enabling efficient sequential integration of task-specific knowledge. Our approach maintains constant memory complexity to the number of models, minimizes interference between tasks through orthogonal projections, and retains the performance of previously merged models through adaptive task vector scaling. Extensive experiments on CLIP-ViT models demonstrate that our method achieves a 5-8% average accuracy improvement while maintaining robust performance in different task orderings. Code is publicly available at https://github.com/tanganke/opcm .

Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Main-stream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile-- with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution-- adding purely random perturbations to the fine-tuned model, can recover the model from harmful behaviors, though it leads to a degradation in the model’s fine-tuning performance. To address the degradation of fine-tuning performance, we further propose \methodname, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. \methodname maintains model's safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.2%, while maintaining fine-tuning performance. As a by-product, we analyze the adaptive perturbation and show that different layers in various LLMs have distinct safety coefficients. Source code available at https://github.com/w-yibo/Panacea.

Large Language Models (LLMs) excel at natural language processing tasks, but their massive size leads to high computational and storage demands. Recent works have sought to reduce their model size through layer-wise structured pruning. However, they tend to ignore retaining the capabilities in the pruned part. In this work, we re-examine structured pruning paradigms and uncover several key limitations: 1) notable performance degradation due to direct layer removal, 2) incompetent linear weighted layer aggregation, and 3) the lack of effective post-training recovery mechanisms. To address these limitations, we propose CoMe, including a progressive layer pruning framework with a Concatenation-based Merging technology and a hierarchical distillation post-training process. Specifically, we introduce a channel sensitivity metric that utilizes activation intensity and weight norms for fine-grained channel selection. Subsequently, we employ a concatenation-based layer merging method to fuse the most critical channels in the adjacent layers, enabling a progressive model size reduction. Finally, we propose a hierarchical distillation protocol, which leverages the correspondences between the original and pruned model layers established during pruning, enabling efficient knowledge transfer. Experiments on seven benchmarks show that CoMe achieves state-of-the-art performance; when pruning 30% of LLaMA-2-7b's parameters, the pruned model retains 83% of its original average accuracy.

Preserving intellectual property (IP) within a pre-trained diffusion model is critical for protecting the model's copyright and preventing unauthorized model deployment. In this regard, model watermarking is a common practice for IP protection that embeds traceable information within models and allows for further verification. Nevertheless, existing watermarking schemes often face challenges due to their vulnerability to fine-tuning, limiting their practical application in general pre-training and fine-tuning paradigms. Inspired by using mode connectivity to analyze model performance between a pair of connected models, we investigate watermark vulnerability by leveraging Linear Mode Connectivity (LMC) as a proxy to analyze the fine-tuning dynamics of watermark performance. Our results show that existing watermarked models tend to converge to sharp minima in the loss landscape, thus making them vulnerable to fine-tuning. To tackle this challenge, we propose **RoMa**, a **Ro**bust **M**odel w**a**termarking scheme that improves the robustness of watermarks against fine-tuning. Specifically, RoMa decomposes watermarking into two components, including *Embedding Functionality*, which preserves reliable watermark detection capability, and *Path-specific Smoothness*, which enhances the smoothness along the watermark-connected path to improve robustness. Extensive experiments on benchmark datasets MS-COCO-2017 and CUB-200-2011 demonstrate that RoMa significantly improves watermark robustness against fine-tuning while maintaining generation quality, outperforming baselines. The code is available at [https://github.com/xiekks/RoMa](https://github.com/xiekks/RoMa).

Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models' parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2\%-5\% gain) and model merging (1\%-3\% gain) methods in achieving balanced LLM alignment.

Model merging integrates multiple expert models with diverse capabilities into a unified framework, facilitating collaborative learning. However, most existing methods assume simultaneous access to all models, which is often impractical in real-world scenarios where models are received sequentially. While some studies have investigated continual model merging (CMM)--which involves sequentially merging multiple models--the challenge of balancing prior knowledge (stability) and incorporating new tasks (plasticity) remains unresolved. This paper, for the first time, formally defines the stability and plasticity of CMM from the perspective of orthogonal projection. Subsequently, we analyze the relationships among the spaces spanned by task data, historical gradients, and accumulated gradients. Building on this, we propose a data-free \textbf{D}ual \textbf{O}rthogonal \textbf{P}rojection (DOP) method, which eliminates data dependence and mitigates interference between the merged model and models for old and new tasks by projecting their parameter differences onto their respective approximate data spaces. Finally, to solve potential conflicts between stability and plasticity, we reformulate DOP as a multi-objective optimization problem and employ a multi-gradient descent algorithm to obtain a Pareto-optimal solution. Extensive experiments across multiple architectures and task configurations validate that our approach significantly outperforms state-of-the-art CMM methods.

In this work, we aim to develop an MLLM that understands and solves questions by learning to create each intermediate step of the reasoning involved till the final answer. To this end, we propose Collective Monte Carlo Tree Search (CoMCTS), a new learning-to-reason method for MLLMs, which introduces the concept of collective learning into ``tree search'' for effective and efficient reasoning-path searching and learning. The core idea of CoMCTS is to leverage collective knowledge from multiple models to collaboratively conjecture, search and identify effective reasoning paths toward correct answers via four iterative operations including Expansion, Simulation and Error Positioning, Backpropagation, and Selection. Using CoMCTS, we construct Mulberry-260k, a multimodal dataset with a tree of rich, explicit and well-defined reasoning nodes for each question. With Mulberry-260k, we perform collective SFT to train our model, Mulberry, a series of MLLMs with o1-like step-by-step Reasoning and Reflection capabilities. Extensive experiments demonstrate the superiority of our proposed methods on various benchmarks. Code is available at https://github.com/HJYao00/Mulberry.

In this work, we aim to incentivize the reasoning ability of Multimodal Large Language Models (MLLMs) via reinforcement learning (RL) and develop an effective approach that mitigates the sparse reward and advantage vanishing issues during RL. To this end, we propose Share-GRPO, a novel RL approach that tackle these issues by exploring and sharing diverse reasoning trajectories over expanded question space. Specifically, Share-GRPO first expands the question space for a given question via data transformation techniques, and then encourages MLLM to effectively explore diverse reasoning trajectories over the expanded question space and shares the discovered reasoning trajectories across the expanded questions during RL. In addition, Share-GRPO also shares reward information during advantage computation, which estimates solution advantages hierarchically across and within question variants, allowing more accurate estimation of relative advantages and improving the stability of policy training. Extensive evaluations over 6 widely-used reasoning benchmarks showcase the superior performance of our method. Code is available at https://github.com/HJYao00/R1-ShareVL.

In this paper, we present two effective policy learning algorithms for multi-agent online coordination(MA-OC) problem. The first one, **MA-SPL**, not only can achieve the optimal $(1-\frac{c}{e})$-approximation guarantee for the MA-OC problem with submodular objectives but also can handle the unexplored $\alpha$-weakly DR-submodular and $(\gamma,\beta)$-weakly submodular scenarios, where $c$ is the curvature of the investigated submodular functions, $\alpha$ denotes the diminishing-return(DR) ratio and the tuple$(\gamma,\beta)$ represents the submodularity ratios. Subsequently, in order to reduce the reliance on the unknown parameters $\alpha,\gamma,\beta$ inherent in the **MA-SPL** algorithm, we then introduce the second online algorithm named **MA-MPL**. This **MA-MPL** algorithm is entirely *parameter-free* and simultaneously can maintain the same approximation ratio as the first **MA-SPL** algorithm. The core of our **MA-SPL** and **MA-MPL** algorithms is a novel continuous-relaxation technique term as policy-based continuous extension. Compared with the well-established multi-linear extension, a notable advantage of this new policy-based continuous extension is its ability to provide a lossless rounding scheme for any set function, thereby enabling us to tackle the challenging weakly submodular objective functions. Finally, extensive simulations are conducted to demonstrate the effectiveness of our proposed algorithms.

The Conditional Sequence Modeling (CSM) paradigm, benefiting from the transformer's powerful distribution modeling capabilities, has demonstrated considerable promise in Reinforcement Learning (RL) tasks. However, much of the work has focused on applying CSM to single online or offline settings, with the general architecture rarely explored. Additionally, existing methods primarily focus on deterministic trajectory modeling, overlooking the randomness of state transitions and the diversity of future trajectory distributions. Fortunately, value-based methods offer a viable solution for CSM, further bridging the potential gap between offline and online RL. In this paper, we propose Value-Guided Decision Transformer (VDT), which leverages value functions to perform advantage-weighting and behavior regularization on the Decision Transformer (DT), guiding the policy toward upper-bound optimal decisions during the offline training phase. In the online tuning phase, VDT further integrates value-based policy improvement with behavior cloning under the CSM architecture through limited interaction and data collection, achieving performance improvement within minimal timesteps. The predictive capability of value functions for future returns is also incorporated into the sampling process. Our method achieves competitive performance on various standard RL benchmarks, providing a feasible solution for developing CSM architectures in general scenarios. Code is available at here.

Constrained hybrid-action reinforcement learning (RL) promises to learn a safe policy within a parameterized action space, which is particularly valuable for safety-critical applications involving discrete-continuous hybrid action spaces. However, existing hybrid-action RL algorithms primarily focus on reward maximization, which faces significant challenges for tasks involving both cost constraints and hybrid action spaces. In this work, we propose a novel Constrained Hybrid-action Policy Optimization algorithm (CHPO) to address the problems of constrained hybrid-action RL. Concretely, we rethink the limitations of hybrid-action RL in handling safe tasks with parameterized action spaces and reframe the objective of constrained hybrid-action RL by introducing the concept of Constrained Parameterized-action Markov Decision Process (CPMDP). Subsequently, we present a constrained hybrid-action policy optimization algorithm to confront the constrained hybrid-action problems and conduct theoretical analyses demonstrating that the CHPO converges to the optimal solution while satisfying safety constraints. Finally, extensive experiments demonstrate that the CHPO achieves competitive performance across multiple experimental tasks.

Federated Learning (FL) enables multiple clients to collaboratively train models without sharing raw data, but is vulnerable to Byzantine attacks and data heterogeneity, which can severely degrade performance. Existing Byzantine-robust approaches tackle data heterogeneity, but incur high computational overhead during gradient aggregation, thereby slowing down the training process. To address this issue, we propose a simple yet effective Federated Normalized Gradients Algorithm (Fed-NGA), which performs aggregation by merely computing the weighted mean of the normalized gradients from each client. This approach yields a favorable time complexity of $\mathcal{O}(pM)$, where $p$ is the model dimension and $M$ is the number of clients. We rigorously prove that Fed-NGA is robust to both Byzantine faults and data heterogeneity. For non-convex loss functions, Fed-NGA achieves convergence to a neighborhood of stationary points under general assumptions, and further attains zero optimality gap under some mild conditions, which is an outcome rarely achieved in existing literature. In both cases, the convergence rate is $\mathcal{O}(1/T^{\frac{1}{2} - \delta})$, where $T$ denotes the number of iterations and $\delta \in (0, 1/2)$. Experimental results on benchmark datasets confirm the superior time efficiency and convergence performance of Fed-NGA over existing methods.

Deep model fusion is an emerging technique that unifies the predictions or parameters of several deep neural networks into a single better-performing model in a cost-effective and data-efficient manner. Although a variety of deep model fusion techniques have been introduced, their evaluations tend to be inconsistent and often inadequate to validate their effectiveness and robustness. We present FusionBench, the first benchmark and a unified library designed specifically for deep model fusion. Our benchmark consists of multiple tasks, each with different settings of models and datasets. This variety allows us to compare fusion methods across different scenarios and model scales. Additionally, FusionBench serves as a unified library for easy implementation and testing of new fusion techniques. FusionBench is open source and actively maintained, with community contributions encouraged.

Large language models (LLMs) exhibit remarkable capabilities in various natural language processing tasks, such as machine translation. However, the large number of LLM parameters incurs significant costs during inference. Previous studies have attempted to train translation-tailored LLMs with moderately sized models by fine-tuning them on the translation data. Nevertheless, when performing translations in zero-shot directions that are absent from the fine-tuning data, the problem of ignoring instructions and thus producing translations in the wrong language (i.e., the off-target translation issue) remains unresolved. In this work, we design a two-stage fine-tuning algorithm to improve the instruction-following ability of translation-tailored LLMs, particularly for maintaining accurate translation directions. We first fine-tune LLMs on the translation data to elicit basic translation capabilities. At the second stage, we construct instruction-conflicting samples by randomly replacing the instructions with the incorrect ones. Then, we introduce an extra unlikelihood loss to reduce the probability assigned to those samples. Experiments on two benchmarks using the LLaMA 2 and LLaMA 3 models, spanning 16 zero-shot directions, demonstrate that, compared to the competitive baseline—translation-finetuned LLaMA, our method could effectively reduce the off-target translation ratio (up to −62.4 percentage points), thus improving translation quality (up to +9.7 bilingual evaluation understudy). Analysis shows that our method can preserve the model’s performance on other tasks, such as supervised translation and general tasks. Code is released at https://github.com/alphadl/LanguageAware_Tuning .

Stochastic gradient descent (SGD) performed in an asynchronous manner plays a crucial role in training large-scale machine learning models. However, the generalization performance of asynchronous delayed SGD, which is an essential metric for assessing machine learning algorithms, has rarely been explored. Existing generalization error bounds are rather pessimistic and cannot reveal the correlation between asynchronous delays and generalization. In this paper, we investigate sharper generalization error bound for SGD with asynchronous delay $\tau$τ. Leveraging the generating function analysis tool, we first establish the average stability of the delayed gradient algorithm. Based on this algorithmic stability, we provide upper bounds on the generalization error of $\widetilde{\mathcal {O}}(\frac{T-\tau }{n\tau })$O˜(T-τnτ) and $\widetilde{\mathcal {O}}(\frac{1}{n})$O˜(1n) for quadratic convex and strongly convex problems, respectively, where $T$T refers to the iteration number and $n$n is the amount of training data. Our theoretical results indicate that asynchronous delays reduce the generalization error of the delayed SGD algorithm. Analogous analysis can be generalized to the random delay setting, and the experimental results validate our theoretical findings.

To address the communication burden issues associated with Federated Learning (FL), Decentralized Federated Learning (DFL) discards the central server and establishes a decentralized communication network, where each client communicates only with neighboring clients. However, existing DFL methods still suffer from two major challenges: local inconsistency and local heterogeneous overfitting, which existing DFL methods have not fundamentally addressed. To tackle these issues, we propose novel DFL algorithms, DFedADMM and its enhanced version DFedADMM-SAM, to improve the performance for DFL. The DFedADMM algorithm employs primal-dual optimization (ADMM) by utilizing dual variables to control the model inconsistency raised from the decentralized heterogeneous data distributions. The DFedADMM-SAM algorithm further improves on DFedADMM by employing a Sharpness-Aware Minimization (SAM) optimizer, which uses gradient perturbations to generate locally flat models and searches for models with uniformly low loss values to mitigate local heterogeneous overfitting. Theoretically, we derive convergence rates of $\mathcal {O}(\frac{1}{\sqrt{KT}}+\frac{1}{KT(1-\psi )^{2}})$O(1KT+1KT(1-ψ)2) and $ \mathcal {O}(\frac{1}{\sqrt{KT}}+\frac{1}{KT(1-\psi )^{2}}+ \frac{1}{T^{3/2}K^{1/2}})$O(1KT+1KT(1-ψ)2+1T3/2K1/2) in the non-convex setting for DFedADMM and DFedADMM-SAM, respectively, where $1 - \psi$1-ψ represents the spectral gap of the gossip matrix. Empirically, extensive experiments on MNIST, CIFAR10, and CIFAR100 datasets demonstrate that our algorithms exhibit superior performance in terms of generalization, convergence speed, and communication overhead compared to existing state-of-the-art (SOTA) optimizers in DFL.

Deep neural networks often suffer from poor generalization due to complex and non-convex loss landscapes. Sharpness-Aware Minimization (SAM) is a popular solution that smooths the loss landscape by minimizing the maximized change of training loss when adding a perturbation to the weight. However, indiscriminate perturbation of SAM on all parameters is suboptimal and results in excessive computation, double the overhead of common optimizers like Stochastic Gradient Descent (SGD). In this paper, we propose Sparse SAM (SSAM), an efficient and effective training scheme that achieves sparse perturbation by a binary mask. To obtain the sparse mask, we provide two solutions based on Fisher information and dynamic sparse training, respectively. We investigate the impact of different masks, including unstructured, structured, and $N$N:$M$M structured patterns, as well as explicit and implicit forms of implementing sparse perturbation. We theoretically prove that SSAM can converge at the same rate as SAM, i.e., $O(\log T/\sqrt{T})$O(logT/T) . Sparse SAM has the potential to accelerate training and smooth the loss landscape effectively. Extensive experimental results on CIFAR and ImageNet-1K confirm that our method is superior to SAM in terms of efficiency, and the performance is preserved or even improved with a perturbation of merely 50% sparsity.

To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), client-level Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharp loss landscape and have poor weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with improved stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. To further reduce the magnitude of random noise while achieving better performance, we propose DP-FedSAM-$\operatorname{top}_{k}$topk by adopting the local update sparsification technique. From the theoretical perspective, we present the convergence analysis to investigate how our algorithms mitigate the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP, the sensitivity analysis of local updates, and generalization analysis. At last, we empirically confirm that our algorithms achieve state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL.

The bounded variance, gradient Lipschitz, and unbiased stochastic gradient are three key assumptions for ensuring the convergence and generalization of stochastic methods, especially in nonconvex scenarios. However, it is important to acknowledge that in practical applications, one or more of these assumptions might be violated, which is the main focus of this paper. In this study, we aim to demonstrate that by incorporating simple gradient normalization with momentum, SGD can effectively guarantee convergence and generalization, even in the presence of unbounded noise, weak gradient Lipschitz, and biased stochastic gradient caused by delays. These results significantly broaden the range of applications for stochastic algorithms, as they relax the previous assumptions and provide more flexibility in real-world scenarios.

Forgetting refers to the loss or deterioration of previously acquired knowledge. While existing surveys on forgetting have primarily focused on continual learning, forgetting is a prevalent phenomenon observed in various other research domains within deep learning. Forgetting manifests in research fields such as generative models due to generator shifts, and federated learning due to heterogeneous data distributions across clients. Addressing forgetting encompasses several challenges, including balancing the retention of old task knowledge with fast learning of new task, managing task interference with conflicting goals, and preventing privacy leakage, etc. Moreover, most existing surveys on continual learning implicitly assume that forgetting is always harmful. In contrast, our survey argues that forgetting is a double-edged sword and can be beneficial and desirable in certain cases, such as privacy-preserving scenarios. By exploring forgetting in a broader context, we present a more nuanced understanding of this phenomenon and highlight its potential advantages. Through this comprehensive survey, we aspire to uncover potential solutions by drawing upon ideas and approaches from various fields that have dealt with forgetting. By examining forgetting beyond its conventional boundaries, we hope to encourage the development of novel strategies for mitigating, harnessing, or even embracing forgetting in real applications.

Constrained optimization problems are pervasive in various fields, and while conventional techniques offer solutions, they often struggle with scalability. Leveraging the power of deep neural networks (DNNs) in optimization, we present a novel learning-based approach, the Constraint Boundary Wandering Framework (CBWF), to address these challenges. Our contributions include introducing a boundary wandering strategy inspired by the active-set method, enhancing equality constraint feasibility, and treating the Lipschitz constant as a learnable parameter. Additionally, we evaluate the regularization term, illustrating that the nonsmooth L2 norm yields superior results. Extensive testing on synthetic datasets and the ACOPT dataset demonstrates CBWF's superiority, outperforming existing deep learning-based solvers in terms of both objective and constraint loss.

The goal of continual learning (CL) is to learn from a series of continuously arriving new tasks without forgetting previously learned old tasks. To avoid catastrophic forgetting of old tasks, orthogonal gradient projection (OGP) based CL methods constrain the gradients of new tasks to be orthogonal to the space spanned by old tasks. This strict gradient constraint will limit the learning ability of new tasks, resulting in lower performance on new tasks. In this paper, we first establish a unified framework for OGP-based CL methods. We then revisit OGP-based CL methods from a new perspective on the loss landscape, where we find that when relaxing projection constraints to improve performance on new tasks, the unflatness of the loss landscape can lead to catastrophic forgetting of old tasks. Based on our findings, we propose a new Dual Flatness-aware OGD framework that optimizes the flatness of the loss landscape from both data and weight levels. Our framework consists of three modules: data and weight perturbation, flatness-aware optimization, and gradient projection. Specifically, we first perform perturbations on the task's data and current model weights to make the task's loss reach the worst-case. Next, we optimize the loss and loss landscape on the original data and the worst-case perturbed data to obtain a flatness-aware gradient. Finally, the flatness-aware gradient will update the network in directions orthogonal to the space spanned by the old tasks. Extensive experiments on four benchmark datasets show that the framework improves the flatness of the loss landscape and performance on new tasks, and achieves state-of-the-art (SOTA) performance on average accuracy across all tasks.

Graph classification has been a prominent problem in graph machine learning fields. This problem has been investigated by leveraging message passing neural networks (MPNNs) to learn powerful graph representations. However, MPNNs extract topological semantics implicitly under label supervision, which could suffer from domain shift and label scarcity in unsupervised domain adaptation settings. In this paper, we propose an effective solution named Dual Variational Semantics Graph Mining (DREAM) for unsupervised graph domain adaptation by combining graph structural semantics from complementary perspectives. Besides a message passing branch to learn implicit semantics, our DREAM trains a path aggregation branch, which can provide explicit high-order structural semantics as a supplement. To train these two branches conjointly, we employ an expectation-maximization (EM) style variational framework for the maximization of likelihood. In the E-step, we fix the message passing branch and construct a graph-of-graph to indicate the geometric correlation between source and target domains, which would be adopted for the optimization of the other branch. In the M-step, we train the message passing branch and update the graph neural networks on the graph-of-graph with the other branch fixed. The alternative optimization improves the collaboration of knowledge from two branches. Extensive experiments on several benchmark datasets validate the superiority of the proposed DREAM compared with various baselines.

Reward finetuning has emerged as a powerful technique for aligning diffusion models with specific downstream objectives or user preferences. However, current approaches suffer from a persistent challenge of reward overoptimization, where models exploit imperfect reward feedback at the expense of overall performance. In this work, we identify three key contributors to overoptimization: (1) a granularity mismatch between the multi-step diffusion process and sparse rewards; (2) a loss of plasticity that limits the model's ability to adapt and generalize; and (3) an overly narrow focus on a single reward objective that neglects complementary performance criteria. Accordingly, we introduce Constrained Diffusion Policy Optimization (CDPO), a novel reinforcement learning framework that addresses reward overoptimization from multiple angles. Firstly, CDPO tackles the granularity mismatch through a temporal policy optimization strategy that delivers step-specific rewards throughout the entire diffusion trajectory, thereby reducing the risk of overfitting to sparse final-step rewards. Then we incorporate a neuron reset strategy that selectively resets overactive neurons in the model, preventing overoptimization induced by plasticity loss. Finally, to avoid overfitting to a narrow reward objective, we integrate constrained reinforcement learning with auxiliary reward objectives serving as explicit constraints, ensuring a balanced optimization across diverse performance metrics.

To address the communication burden and privacy concerns associated with the centralized server in Federated Learning (FL), Decentralized Federated Learning (DFL) has emerged, which discards the server with a peer-to-peer (P2P) communication framework, significantly expanding the application scenarios of FL. However, most existing DFL algorithms are based on symmetric topologies, such as ring and grid topology, which can easily lead to deadlocks and are susceptible to the impact of network link quality in practice. To address these issues, we propose DFedSGPSM, a transitional framework that converts symmetric DFL optimizers into asymmetric variants. By adopting the Push-Sum protocol in asymmetric network topologies, our framework successfully circumvents the deadlock and link-quality issues prevalent in symmetric configurations. To further validate the effectiveness of our algorithm framework, we integrate the local momentum (in DFedAvgM) and SAM (in DFedSAM) from existing symmetric DFL optimizer into DFedSGPSM to accelerate training and pursue smooth local minimum, which enables existing symmetric DFL optimizers to be seamlessly integrated into asymmetric DFL. Theoretical analysis proves that DFedSGPSM achieves a linear speedup rate of$\mathcal{O}(\frac{1}{\sqrt{nT}})$in the non-convex setting. This analysis also reveals crucial issues such as tighter upper bounds achieved with improved topological connectivity. Empirically, extensive experiments conducted on the MNIST, CIFAR10&100 datasets demonstrate the superior performance of our proposed algorithm compared to several existing SOTA optimizers in terms of generalization.

The impressive performance of transformer models has sparked the deployment of intelligent applications on resource-constrained edge devices. However, ensuring high-quality service for real-time edge systems is a significant challenge due to the considerable computational demands and resource requirements of these models. Existing strategies typically either offload transformer computations to other devices or directly deploy compressed models on individual edge devices. These strategies, however, result in either considerable communication overhead or suboptimal trade-offs between accuracy and efficiency. To tackle these challenges, we propose a collaborative inference system for general transformer models, termed CoFormer. The central idea behind CoFormer is to exploit the divisibility and integrability of transformer. An off-the-shelf large transformer can be decomposed into multiple smaller models for distributed inference, and their intermediate results are aggregated to generate the final output. We formulate an optimization problem to minimize both inference latency and accuracy degradation under heterogeneous hardware constraints. DeBo algorithm is proposed to first solve the optimization problem to derive the decomposition policy, and then progressively calibrate decomposed models to restore performance. We demonstrate the capability to support a wide range of transformer models on heterogeneous edge devices, achieving up to 3.1× inference speedup with large transformer models. Notably, CoFormer enables the efficient inference of GPT2-XL with 1.6 billion parameters on edge devices, reducing memory requirements by 76.3%. CoFormer can also reduce energy consumption by approximately 40% while maintaining satisfactory inference performance.

The goal of single domain generalization is to use data from a single domain (source domain) to train a model, which is then deployed over several unknown domains for testing (target domains). This study introduces a practical approach diverging from traditional DG, which typically relies on multiple source domains. We focus on Single Long-Tailed Domain Generalization, which refers to a scenario in the context of long-tail distribution, where although minority classes may have fewer samples in a single domain, these minority classes could become more prevalent and dominant in other domains. We introduce the Graph Convolutional Mixture-of-Experts Learners Network for Long-Tailed Domain Generalization (GCML) as a solution to this problem. Our approach presents two novel tactics. Initially, we utilize an expert learning technique that is skill-diverse. In order to properly manage the unknown target domain, this entails training multiple specialists inside a single long-tailed source domain and combining their knowledge. Then, we use a graph convolutional network to facilitate domain generalization, leveraging joint data structure modeling to learn more domain-invariant feature. Experiments conducted on four established benchmarks reveal that our GCML algorithm outperforms contemporary domain generalization techniques, demonstrating its efficacy in this complex task.

Harmful fine-tuning attack poses a serious threat to the online fine-tuning service. Vaccine, a recent alignment-stage defense, applies uniform perturbation to all layers of embedding to make the model robust to the simulated embedding drift. However, applying layer-wise uniform perturbation may lead to excess perturbations for some particular non-safety-critical layers, resulting in defense performance degradation and unnecessary memory consumption. To address this limitation, we propose a Targeted Vaccine (T-Vaccine), a memory-efficient safety alignment method that applies perturbation to only selected layers of the model. T-Vaccine follows two core steps: First, it uses the harmful gradient norm as a statistical metric to identify the safety-critical layers. Second, instead of applying uniform perturbation across all layers, T-Vaccine only applies perturbation to the safety-critical layers while keeping other layers frozen during training. Results show that T-Vaccine outperforms Vaccine in terms of both defense effectiveness and resource efficiency. Comparison with other defense baselines, e.g., RepNoise and TAR also demonstrate the superiority of T-Vaccine. Notably, T-Vaccine is the first defense that enables a fine-tuning-based alignment method for 7B pre-trained models trained on consumer GPUs with limited memory (e.g., RTX 4090).

Recent advancements in vision transformers (ViTs) have demonstrated that larger models often achieve superior performance. However, training these models remains computationally intensive and costly. To address this challenge, we introduce ScaleNet, an efficient approach for scaling ViT models. Unlike conventional training from scratch, ScaleNet facilitates rapid model expansion with negligible increases in parameters, building on existing pretrained models. This offers a cost-effective solution for scaling up ViTs. Specifically, ScaleNet achieves model expansion by inserting additional layers into pretrained ViTs, utilizing layer-wise weight sharing to maintain parameters efficiency. Each added layer shares its parameter tensor with a corresponding layer from the pretrained model. To mitigate potential performance degradation due to shared weights, ScaleNet introduces a small set of adjustment parameters for each layer. These adjustment parameters are implemented through parallel adapter modules, ensuring that each instance of the shared parameter tensor remains distinct and optimized for its specific function. Experiments on the ImageNet-1K dataset demonstrate that ScaleNet enables efficient expansion of ViT models. With a $2\times $ depth-scaled DeiT-Base model, ScaleNet achieves a 7.42% accuracy improvement over training from scratch while requiring only one-third of the training epochs, highlighting its efficiency in scaling ViTs. Beyond image classification, our method shows significant potential for application in downstream vision areas, as evidenced by the validation in object detection task.

Recently, MLP-based architectures have achieved competitive performance with convolutional neural networks (CNNs) and vision transformers (ViTs) across various vision tasks. However, most MLP-based methods introduce local feature interactions to facilitate direct adaptation to downstream tasks, thereby lacking the ability to capture global visual dependencies and multi-scale context, ultimately resulting in unsatisfactory performance on dense prediction. This paper proposes a competitive and effective MLP-based architecture called Pyramid Fusion MLP (PFMLP) to address the above limitation. Specifically, each block in PFMLP introduces multi-scale pooling and fully connected layers to generate feature pyramids, which are subsequently fused using up-sample layers and an additional fully connected layer. Employing different down-sample rates allows us to obtain diverse receptive fields, enabling the model to simultaneously capture long-range dependencies and fine-grained cues, thereby exploiting the potential of global context information and enhancing the spatial representation power of the model. Our PFMLP is the first lightweight MLP to obtain comparable results with state-of-the-art CNNs and ViTs on the ImageNet-1K benchmark.With larger FLOPs, it exceeds state-of-the-art CNNs, ViTs, and MLPs under similar computational complexity. Furthermore, experiments in object detection, instance segmentation, and semantic segmentation demonstrate that the visual representation acquired from PFMLP can be seamlessly transferred to downstream tasks, producing competitive results. All materials contain the training codes and logs are released at https://github.com/huangqiuyu/PFMLP.

Unsupervised efficient domain adaptive retrieval aims to transfer knowledge from a labeled source domain to an unlabeled target domain, while maintaining low storage cost and high retrieval efficiency. However, existing methods typically fail to address potential noise in the target domain, and directly align high-level features across domains, thus resulting in suboptimal retrieval performance. To address these challenges, we propose a novel Cross-Domain Diffusion with Progressive Alignment method (COUPLE). This approach revisits unsupervised efficient domain adaptive retrieval from a graph diffusion perspective, simulating cross-domain adaptation dynamics to achieve a stable target domain adaptation process. First, we construct a cross-domain relationship graph and leverage noise-robust graph flow diffusion to simulate the transfer dynamics from the source domain to the target domain, identifying lower noise clusters. We then leverage the graph diffusion results for discriminative hash code learning, effectively learning from the target domain while reducing the negative impact of noise. Furthermore, we employ a hierarchical Mixup operation for progressive domain alignment, which is performed along the cross-domain random walk paths. Utilizing target domain discriminative hash learning and progressive domain alignment, COUPLE enables effective domain adaptive hash learning. Extensive experiments demonstrate COUPLE's effectiveness on competitive benchmarks.

In a real federated learning (FL) system, communication overhead for passing model parameters between the clients and the parameter server (PS) is often a bottleneck. Hierarchical federated learning (HFL) that poses multiple edge servers (ESs) between clients and the PS can partially alleviate communication pressure but still needs the aggregation of model parameters from multiple ESs at the PS. To further reduce communication overhead, we remove the central PS, so that each iteration only completes model training by transmitting the global model between two adjacent ES. We call this serial learning method Sequential FL (SFL). For the first time, we introduced SFL into HFL and proposed a novel algorithm adapted to this combined framework, called Fed-CHS. Convergence results are derived for strongly convex and non-convex loss functions under various data heterogeneity setups, which show comparable convergence performance with the algorithms for HFL or SFL solely. Experimental results provide evidence of the superiority of our proposed Fed-CHS on both communication overhead saving and test accuracy over baseline methods.

Federated learning (FL) aims to collaboratively learn a model by using the data from multiple users under privacy constraints. In this article, we study the multilabel classification (MLC) problem under the FL setting, where trivial solution and extremely poor performance may be obtained, especially when only positive data with respect to a single class label is provided for each client. This issue can be addressed by adding a specially designed regularizer on the server side. Although effective sometimes, the label correlations are simply ignored and thus suboptimal performance may be obtained. Besides, it is expensive and unsafe to exchange user's private embeddings between server and clients frequently, especially when training model in the contrastive way. To remedy these drawbacks, we propose a novel and generic method termed federated averaging (FedAvg) by exploring label correlations (FedALCs). Specifically, FedALC estimates the label correlations in the class embedding learning for different label pairs and utilizes it to improve the model training. To further improve the safety and also reduce the communication overhead, we propose a variant to learn fixed class embedding for each client, so that the server and clients only need to exchange class embeddings once. Extensive experiments on multiple popular datasets demonstrate that our FedALC can significantly outperform the existing counterparts.

Artificial neural networks, especially recent diffusion-based models, have shown remarkable superiority in gaming, control, and QA systems, where the training tasks' datasets are usually static. However, in real-world applications, such as robotic control of reinforcement learning (RL), the tasks are changing, and new tasks arise in a sequential order. This situation poses the new challenge of plasticity-stability tradeoff for training an agent who can adapt to task changes and retain acquired knowledge. In view of this, we propose a rehearsal-based continual diffusion model, called continual diffuser (CoD), to endow the diffuser with the capabilities of quick adaptation (plasticity) and lasting retention (stability). Specifically, we first construct an offline benchmark that contains 90 tasks from multiple domains. Then, we train the CoD on each task with sequential modeling and conditional generation for making decisions. Next, we preserve a small portion of previous datasets as the rehearsal buffer and replay it to retain the acquired knowledge. Extensive experiments on a series of tasks show that CoD can achieve a promising plasticity-stability tradeoff and outperform existing diffusion-based methods and other representative baselines on most tasks. The source code is available at https://github.com/JF-Hu/Continual_Diffuser.

Federated learning (FL) is a distributed machine learning paradigm that enables heterogeneous devices to train a model collaboratively. Recognizing communication as a bottleneck in FL, existing communication-efficient solutions, e.g., HeteroFL and LotteryFL, etc., utilize gradient sparsification to reduce communication costs. However, existing solutions fail to address the dynamic bandwidth issue in which the bandwidth of each client is constantly changing throughout the training process. In this article, we propose AdaptiveFL, a communication-adaptive FL framework, considering the dynamic constraints of bandwidth. The design of AdaptiveFL follows two key steps: 1) in each round, each device selects a best-fit sub-model for communication per currently available bandwidth; and 2) to guarantee the performance of each sub-model sent under dynamic bandwidth constraints, AdaptiveFL employs a local training method that enables each device to train a "tailorable" local model, which can be tailored to any sparsity with competitive accuracy. We compare AdaptiveFL with several communication-efficient SOTA methods and demonstrate that AdaptiveFL outperforms other baselines by a large margin.

Deep reinforcement learning (RL) has shown remarkable success in specific offline decision-making scenarios, yet its theoretical guarantees are still under development. Existing works on offline RL theory primarily emphasize a few trivial settings, such as linear MDP or general function approximation with strong assumptions and independent data, which lack guidance for practical use. The coupling of deep learning and Bellman residuals makes this problem challenging, in addition to the difficulty of data dependence. In this paper, we establish a non-asymptotic estimation error of pessimistic offline RL using general neural network approximation with C-mixing data regarding the structure of networks, the dimension of datasets, and the concentrability of data coverage, under mild assumptions. Our result shows that the estimation error consists of two parts: the first converges to zero at a desired rate on the sample size with partially controllable concentrability, and the second becomes negligible if the residual constraint is tight. This result demonstrates the explicit efficiency of deep adversarial offline RL frameworks. We utilize the empirical process tool for C-mixing sequences and the neural network approximation theory for the Holder class to achieve this. We also develop methods to bound the Bellman estimation error caused by function approximation with empirical Bellman constraint perturbations. Additionally, we present a result that lessens the curse of dimensionality using data with low intrinsic dimensionality and function classes with low complexity. Our estimation provides valuable insights into the development of deep offline RL and guidance for algorithm model design.

Knowledge distillation (KD) is a common approach to compress a teacher model to reduce its inference cost and memory footprint, by training a smaller student model.However, in the context of autoregressive language models (LMs), we empirically find that larger teachers might dramatically result in a poorer student.In response to this problem, we conduct a series of analyses and reveal that different tokens have different teaching modes, neglecting which will lead to performance degradation.Motivated by this, we propose a simple yet effective adaptive teaching approach (ATKD) to improve the KD.The core of ATKD is to reduce rote learning and make teaching more diverse and flexible.Extensive experiments on 8 LM tasks show that, with the help of ATKD, various baseline KD methods can achieve consistent and significant performance gains (up to +3.04% average score) across all model types and sizes.More encouragingly, ATKD can improve the student model generalization effectively.

Multi-constraint offline reinforcement learning (RL) promises to learn policies that satisfy both cumulative and state- wise costs from offline datasets. This arrangement provides an effective approach for the widespread appli-cation of RL in high-risk scenarios where both cumulative and state-wise costs need to be considered simulta-neously. However, previously constrained offline RL algorithms are primarily designed to handle single-constraint problems related to cumulative cost, which faces challenges when addressing multi-constraint tasks that involve both cumulative and state-wise costs. In this work, we pro-pose a novel Primal policy Optimization with Conservative Estimation algorithm (POCE) to address the problem of multi-constraint offline RL. Concretely, we reframe the ob-jective of multi-constraint offline RL by introducing the con-cept of Maximum Markov Decision Processes (MMDP). Subsequently, we present a primal policy optimization al-gorithm to confront the multi-constraint problems, which improves the stability and convergence speed of model training. Furthermore, we propose a conditional Bell-man operator to estimate cumulative and state-wise Q-values, reducing the extrapolation error caused by out-of-distribution (OOD) actions. Finally, extensive experiments demonstrate that the POCE algorithm achieves competitive performance across multiple experimental tasks, particu-larly outperforming baseline algorithms in terms of safety. Our code is available at github. POCE.

Recently, non-transferable learning (NTL) was proposed to restrict models' generalization toward the target do-main(s), which serves as state-of-the-art solutions for in-tellectual property (IP) protection. However, the robustness of the established “transferability barrier” for degrading the target domain performance has not been well studied. In this paper, we first show that the generalization performance of NTL models is widely impaired on third-party do-mains (i.e., the unseen domain in the NTL training stage). We explore the impairment patterns and find that: due to the dominant generalization of non-transferable task, NTL models tend to make target-domain-consistent predictions on third-party domains, even though only a slight distribution shift from the third-party domain to the source domain. Motivated by these findings, we uncover the potential risks of NTL by proposing a simple but effective method (dubbed as TransNTL) to recover the target domain performance with few source domain data. Specifically, by performing a group of different perturbations on the few source domain data, we obtain diverse third-party domains that evoke the same impairment patterns as the unavailable target domain. Then, we fine-tune the NTL model under an impairment-repair self-distillation framework, where the source-domain predictions are used to teach the model itself how to pre-dict on third-party domains, thus repairing the impaired generalization. Empirically, experiments on standard NTL benchmarks show that the proposed TransNTL reaches up to ∼ 72% target-domain improvements by using only 10% source domain data. Finally, we also explore a feasible de-fense method and empirically demonstrate its effectiveness.

Personalized Federated Learning (PFL) is proposed to find the greatest personalized models for each client. To avoid the central failure and communication bottleneck in the server-based FL, we concentrate on the Decentralized Personalized Federated Learning (DPFL) that performs distributed model training in a Peer-to-Peer (P2P) manner. Most personalized works in DPFL are based on undi-rected and symmetric topologies, however, the data, computation and communication resources heterogeneity result in large variances in the personalized models, which lead the undirected aggregation to suboptimal personalized per-formance and unguaranteed convergence. To address these issues, we propose a directed collaboration DPFL framework by incorporating stochastic gradient push and partial model personalized, called Decentralized Federated Partial Gradient Push (DFedPGP). It personalizes the linear clas-sifier in the modern deep model to customize the local solution and learns a consensus representation in a fully de-centralized manner. Clients only share gradients with a subset of neighbors based on the directed and asymmetric topologies, which guarantees flexible choices for resource efficiency and better convergence. Theoretically, we show that the proposed DFedPGP achieves a superior conver-gence rate of O (1/√T) in the general non-convex setting, and prove the tighter connectivity among clients will speed up the convergence. The proposed method achieves state-of-the-art (SOTA) accuracy in both data and computation heterogeneity scenarios, demonstrating the efficiency of the directed collaboration and partial gradient push.

Data-Free Meta-Learning (DFML) aims to extract knowledge from a collection of pre-trained models without requiring the original data, presenting practical benefits in contexts constrained by data privacy concerns. Current DFML methods primarily focus on the data recovery from these pre-trained models. However, they suffer from slow recovery speed and overlook gaps inherent in heterogeneous pre-trained models. In response to these challenges, we introduce the Faster and Better Data-Free Meta-Learning (FREE) framework, which contains: (i) a meta-generator for rapidly recovering training tasks from pre-trained models; and (ii) a meta-learner for generalizing to new unseen tasks. Specifically, within the module Faster Inversion via Meta-Generator, each pre-trained model is perceived as a distinct task. The meta-generator can rapidly adapt to a specific task in just five steps, significantly accelerating the data recovery. Furthermore, we propose Better Generalization via Meta-Learner and introduce an implicit gradient alignment algorithm to optimize the meta-learner. This is achieved as aligned gradient directions alleviate potential conflicts among tasks from heterogeneous pre-trained models. Empirical experiments on multiple benchmarks affirm the superiority of our approach, marking a notable speed-up (20x) and performance enhancement (1.42% ~ 4.78%) in comparison to the state-of-the-art.

While large language models (LLMs) excel in a simulated world of texts, they struggle to interact with the more realistic world without perceptions of other modalities such as visual or audio signals. Although vision-language models (VLMs) integrate LLM modules (1) aligned with static image features, and (2) may possess prior knowledge of world dynamics (as demonstrated in the text world), they have not been trained in an embodied visual world and thus cannot align with its dynamics. On the other hand, training an embodied agent in a noisy visual world without expert guidance is often chal-lenging and inefficient. In this paper, we train a VLM agent living in a visual world using an LLM agent excelling in a parallel text world. Specifically, we distill LLM's reflection outcomes (improved actions by analyzing mistakes) in a text world's tasks to finetune the VLM on the same tasks of the visual world, resulting in an Embodied Multi-Modal Agent (EMMA) quickly adapting to the visual world dy-namics. Such cross-modality imitation learning between the two parallel worlds is achieved by a novel DAgger-DPO algorithm, enabling EMMA to generalize to a broad scope of new tasks without any further guidance from the LLM expert. Extensive evaluations on the ALFWorld benchmark's diverse tasks highlight EMMA's superior performance to SOTA VLM-based agents, e.g., 20%-70% improvement in the success rate.

Fine-tuning is the process of extending the training of pre-trained models on specific target tasks, thereby significantly enhancing their performance across various applications. However, fine-tuning often demands large memory consumption, posing a challenge for low-memory devices that some previous memory-efficient fine-tuning methods attempted to mitigate by pruning activations for gradient computation, albeit at the cost of significant computational overhead from the pruning processes during training. To address these challenges, we introduce PreBackRazor; a novel activation pruning scheme offering both computational and memory efficiency through a sparsified back-propagation strategy, which judiciously avoids unnecessary activation pruning and storage and gradient computation. Before activation pruning, our approach samples a probability of selecting a portion of parameters to freeze, utilizing a bandit method for updates to prioritize impactful gradients on convergence. During the feed-forward pass, each model layer adjusts adaptively based on parameter activation status, obviating the need for sparsification and storage of redundant activations for subsequent backpropagation. Benchmarking on fine-tuning foundation models, our approach maintains baseline accuracy across diverse tasks, yielding over 20% speedup and around 10% memory reduction. Moreover, integrating with an advanced CUDA kernel achieves up to 60% speedup without extra memory costs or accuracy loss, significantly enhancing the efficiency of fine-tuning foundation models on memory-constrained devices.

The Mixture of Experts (MoE) has emerged as a highly successful technique in deep learning, based on the principle of divide-and-conquer to maximize model capacity without significant additional computational cost. Even in the era of large-scale language models (LLMs), MoE continues to play a crucial role, as some researchers have indicated that GPT-4 adopts the MoE structure to ensure diverse inference results. However, MoE is susceptible to performance degeneracy, particularly evident in the issues of imbalance and homogeneous representation among experts. While previous studies have extensively addressed the problem of imbalance, the challenge of homogeneous representation remains unresolved. In this study, we shed light on the homogeneous representation problem, wherein experts in the MoE fail to specialize and lack diversity, leading to frustratingly high similarities in their representations (up to 99% in a well-performed MoE model). This problem restricts the expressive power of the MoE and, we argue, contradicts its original intention. To tackle this issue, we propose a straightforward yet highly effective solution: OMoE, an orthogonal expert optimizer. Additionally, we introduce an alternating training strategy that encourages each expert to update in a direction orthogonal to the subspace spanned by other experts. Our algorithm facilitates MoE training in two key ways: firstly, it explicitly enhances representation diversity, and secondly, it implicitly fosters interaction between experts during orthogonal weights computation. Through extensive experiments, we demonstrate that our proposed optimization algorithm significantly improves the performance of fine-tuning the MoE model on the GLUE benchmark, SuperGLUE benchmark, question-answering task, and name entity recognition tasks.

Abhinav Bandari, Lu Yin, Cheng-Yu Hsieh, Ajay Kumar Jaiswal, Tianlong Chen, Li Shen, Ranjay Krishna, Shiwei Liu. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024.

Non-transferable learning (NTL) aims to restrict the generalization of models toward the target domain(s). To this end, existing works learn non-transferable representations by reducing statistical dependence between the source and target domain. However, such statistical methods essentially neglect to distinguish between *styles* and *contents*, leading them to inadvertently fit (i) spurious correlation between *styles* and *labels*, and (ii) fake independence between *contents* and *labels*. Consequently, their performance will be limited when natural distribution shifts occur or malicious intervention is imposed. In this paper, we propose a novel method (dubbed as H-NTL) to understand and advance the NTL problem by introducing a causal model to separately model *content* and *style* as two latent factors, based on which we disentangle and harness them as guidances for learning non-transferable representations with intrinsically causal relationships. Specifically, to avoid fitting spurious correlation and fake independence, we propose a variational inference framework to disentangle the naturally mixed *content factors* and *style factors* under our causal model. Subsequently, based on dual-path knowledge distillation, we harness the disentangled two *factors* as guidances for non-transferable representation learning: (i) we constraint the source domain representations to fit *content factors* (which are the intrinsic cause of *labels*), and (ii) we enforce that the target domain representations fit *style factors* which barely can predict labels. As a result, the learned feature representations follow optimal untransferability toward the target domain and minimal negative influence on the source domain, thus enabling better NTL performance. Empirically, the proposed H-NTL significantly outperforms competing methods by a large margin.

In numerous artificial intelligence applications, the collaborative efforts of multiple intelligent agents are imperative for the successful attainment of target objectives. To enhance coordination among these agents, a distributed communication framework is often employed. However, information sharing among all agents proves to be resource-intensive, while the adoption of a manually pre-defined communication architecture imposes limitations on inter-agent communication, thereby constraining the potential for collaborative efforts. In this study, we introduce a novel approach wherein we conceptualize the communication architecture among agents as a learnable graph. We formulate this problem as the task of determining the communication graph while enabling the architecture parameters to update normally, thus necessitating a bi-level optimization process. Utilizing continuous relaxation of the graph representation and incorporating attention units, our proposed approach, CommFormer, efficiently optimizes the communication graph and concurrently refines architectural parameters through gradient descent in an end-to-end manner. Extensive experiments on a variety of cooperative tasks substantiate the robustness of our model across diverse cooperative scenarios, where agents are able to develop more coordinated and sophisticated strategies regardless of changes in the number of agents.

Plasticity, the ability of a neural network to evolve with new data, is crucial for high-performance and sample-efficient visual reinforcement learning (VRL). Although methods like resetting and regularization can potentially mitigate plasticity loss, the influences of various components within the VRL framework on the agent's plasticity are still poorly understood. In this work, we conduct a systematic empirical exploration focusing on three primary underexplored facets and derive the following insightful conclusions: (1) data augmentation is essential in maintaining plasticity; (2) the critic's plasticity loss serves as the principal bottleneck impeding efficient training; and (3) without timely intervention to recover critic's plasticity in the early stages, its loss becomes catastrophic. These insights suggest a novel strategy to address the high replay ratio (RR) dilemma, where exacerbated plasticity loss hinders the potential improvements of sample efficiency brought by increased reuse frequency. Rather than setting a static RR for the entire training process, we propose Adaptive RR, which dynamically adjusts the RR based on the critic’s plasticity level. Extensive evaluations indicate that Adaptive RR not only avoids catastrophic plasticity loss in the early stages but also benefits from more frequent reuse in later phases, resulting in superior sample efficiency.

Large pre-trained models have enabled significant advances in machine learning and served as foundation components. Model fusion methods, such as task arithmetic, have been proven to be powerful and scalable to incorporate fine-tuned weights from different tasks into a multi-task model. However, efficiently fine-tuning large pre-trained models on multiple downstream tasks remains challenging, leading to inefficient multi-task model fusion. In this work, we propose a novel method to improve multi-task fusion for parameter-efficient fine-tuning techniques like LoRA fine-tuning. Specifically, our approach partially linearizes only the adapter modules and applies task arithmetic over the linearized adapters. This allows us to leverage the the advantages of model fusion over linearized fine-tuning, while still performing fine-tuning and inference efficiently. We demonstrate that our partial linearization technique enables a more effective fusion of multiple tasks into a single model, outperforming standard adapter tuning and task arithmetic alone. Experimental results demonstrate the capabilities of our proposed partial linearization technique to effectively construct unified multi-task models via the fusion of fine-tuned task vectors. We evaluate performance over an increasing number of tasks and find that our approach outperforms standard parameter-efficient fine-tuning techniques. The results highlight the benefits of partial linearization for scalable and efficient multi-task model fusion.

Continual Learning (CL) focuses on learning from dynamic and changing data distributions while retaining previously acquired knowledge. Various methods have been developed to address the challenge of catastrophic forgetting, including regularization-based, Bayesian-based, and memory-replay-based techniques. However, these methods lack a unified framework and common terminology for describing their approaches. This research aims to bridge this gap by introducing a comprehensive and overarching framework that encompasses and reconciles these existing methodologies. Notably, this new framework is capable of encompassing established CL approaches as special instances within a unified and general optimization objective. An intriguing finding is that despite their diverse origins, these methods share common mathematical structures. This observation highlights the compatibility of these seemingly distinct techniques, revealing their interconnectedness through a shared underlying optimization objective. Moreover, the proposed general framework introduces an innovative concept called *refresh learning*, specifically designed to enhance the CL performance. This novel approach draws inspiration from neuroscience, where the human brain often sheds outdated information to improve the retention of crucial knowledge and facilitate the acquisition of new information. In essence, *refresh learning* operates by initially unlearning current data and subsequently relearning it. It serves as a versatile plug-in that seamlessly integrates with existing CL methods, offering an adaptable and effective enhancement to the learning process. Extensive experiments on CL benchmarks and theoretical analysis demonstrate the effectiveness of the proposed *refresh learning*.

Multi-task learning (MTL) aims to empower a model to tackle multiple tasks simultaneously. A recent development known as task arithmetic has revealed that several models, each fine-tuned for distinct tasks, can be directly merged into a single model to execute MTL without necessitating a retraining process using the initial training data. Nevertheless, this direct addition of models often leads to a significant deterioration in the overall performance of the merged model. This decline occurs due to potential conflicts and intricate correlations among the multiple tasks. Consequently, the challenge emerges of how to merge pre-trained models more effectively without using their original training data. This paper introduces an innovative technique called Adaptive Model Merging (AdaMerging). This approach aims to autonomously learn the coefficients for model merging, either in a task-wise or layer-wise manner, without relying on the original training data. Specifically, our AdaMerging method operates as an automatic, unsupervised task arithmetic scheme. It leverages entropy minimization on unlabeled test samples from the multi-task setup as a surrogate objective function to iteratively refine the merging coefficients of the multiple models. Our experimental findings across eight tasks demonstrate the efficacy of the AdaMerging scheme we put forth. Compared to the current state-of-the-art (SOTA) task arithmetic merging scheme, AdaMerging showcases a remarkable 11\% improvement in performance. Notably, AdaMerging also exhibits superior generalization capabilities when applied to unseen downstream tasks. Furthermore, it displays a significantly enhanced robustness to data distribution shifts that may occur during the testing phase.

Recently, numerous graph neural network methods have been developed to tackle domain shifts in graph data. However, these methods presuppose that unlabeled target graphs belong to categories previously seen in the source domain. This assumption could not hold true for in-the-wild target graphs. In this paper, we delve deeper to explore a more realistic problem open-set graph domain adaptation. Our objective is to not only identify target graphs from new categories but also accurately classify remaining target graphs into their respective categories under domain shift and label scarcity. To solve this challenging problem, we introduce a new method named Dual Structured Exploration with Mixup (DREAM). DREAM incorporates a graph-level representation learning branch as well as a subgraph-enhanced branch, which jointly explores graph topological structures from both global and local viewpoints. To maximize the use of unlabeled target graphs, we train these two branches simultaneously using posterior regularization to enhance their inter-module consistency. To accommodate the open-set setting, we amalgamate dissimilar samples to generate virtual unknown samples belonging to novel classes. Moreover, to alleviate domain shift, we establish a k nearest neighbor-based graph-of-graphs and blend multiple neighbors of each sample to produce cross-domain virtual samples for inter-domain consistency learning. Extensive experiments validate the effectiveness of the proposed DREAM in comparison to various state-of-the-art approaches in different settings.

The purpose of offline multi-task reinforcement learning (MTRL) is to develop a unified policy applicable to diverse tasks without the need for online environmental interaction. Recent advancements approach this through sequence modeling, leveraging the Transformer architecture’s scalability and the benefits of parameter sharing to exploit task similarities. However, variations in task content and complexity pose significant challenges in policy formulation, necessitating judicious parameter sharing and management of conflicting gradients for optimal policy performance. In this work, we introduce the Harmony Multi-Task Decision Transformer (HarmoDT), a novel solution designed to identify an optimal harmony subspace of parameters for each task. We approach this as a bi-level optimization problem, employing a meta-learning framework that leverages gradient-based techniques. The upper level of this framework is dedicated to learning a task-specific mask that delineates the harmony subspace, while the inner level focuses on updating parameters to enhance the overall performance of the unified policy. Empirical evaluations on a series of benchmarks demonstrate the superiority of HarmoDT, verifying the effectiveness of our approach.

Recent advancements in offline reinforcement learning (RL) have underscored the capabilities of Conditional Sequence Modeling (CSM), a paradigm that learns the action distribution based on history trajectory and target returns for each state. However, these methods often struggle with stitching together optimal trajectories from sub-optimal ones due to the inconsistency between the sampled returns within individual trajectories and the optimal returns across multiple trajectories. Fortunately, Dynamic Programming (DP) methods offer a solution by leveraging a value function to approximate optimal future returns for each state, while these techniques are prone to unstable learning behaviors, particularly in long-horizon and sparse-reward scenarios. Building upon these insights, we propose the Q-value regularized Transformer (QT), which combines the trajectory modeling ability of the Transformer with the predictability of optimal future returns from DP methods. QT learns an action-value function and integrates a term maximizing action-values into the training loss of CSM, which aims to seek optimal actions that align closely with the behavior policy. Empirical evaluations on D4RL benchmark datasets demonstrate the superiority of QT over traditional DP and CSM methods, highlighting the potential of QT to enhance the state-of-the-art in offline RL.

Model inversion, which aims to reconstruct the original training data from pre-trained discriminative models, is especially useful when the original training data is unavailable due to privacy, usage rights, or size constraints. However, existing dense inversion methods attempt to reconstruct the entire image area, making them extremely inefficient when inverting high-resolution images from large-scale Vision Transformers (ViTs). We further identify two underlying causes of this inefficiency: the redundant inversion of noisy backgrounds and the unintended inversion of spurious correlations—a phenomenon we term “hallucination” in model inversion. To address these limitations, we propose a novel sparse model inversion strategy, as a plug-and-play extension to speed up existing dense inversion methods with no need for modifying their original loss functions. Specifically, we selectively invert semantic foregrounds while stopping the inversion of noisy backgrounds and potential spurious correlations. Through both theoretical and empirical studies, we validate the efficacy of our approach in achieving significant inversion acceleration (up to $\times$3.79) while maintaining comparable or even enhanced downstream performance in data-free model quantization and data-free knowledge transfer. Code is available at https://github.com/Egg-Hu/SMI.

Merging various task-specific Transformer-based vision models trained on different tasks into a single unified model can execute all the tasks concurrently. Previous methods, exemplified by task arithmetic, have been proven to be both effective and scalable. Existing methods have primarily focused on seeking a static optimal solution within the original model parameter space. A notable challenge is mitigating the interference between parameters of different models, which can substantially deteriorate performance. In this paper, we propose to merge most of the parameters while upscaling the MLP of the Transformer layers to a weight-ensembling mixture of experts (MoE) module, which can dynamically integrate shared and task-specific knowledge based on the input, thereby providing a more flexible solution that can adapt to the specific needs of each instance. Our key insight is that by identifying and separating shared knowledge and task-specific knowledge, and then dynamically integrating them, we can mitigate the parameter interference problem to a great extent. We conduct the conventional multi-task model merging experiments and evaluate the generalization and robustness of our method. The results demonstrate the effectiveness of our method and provide a comprehensive understanding of our method. The code is available at https://github.com/tanganke/weight-ensembling_MoE

Stochastic weight averaging (SWA) method has empirically proven its advantages compared to stochastic gradient descent (SGD). Despite it is widespread used, theoretical investigations have been limited, particularly in scenarios beyond the ideal setting of convex and sampling with replacement. However, non-convex cases and sampling without replacement are very practical in real-world applications. The main challenges under the above settings are two-folds: (i) All the historical gradient information introduced by SWA is considered, while the analysis of SGD using the tool of uniform stability requires only to bound the current gradient. (ii) The $(1+\alpha\beta)$-expansion property causes the boundary of each gradient step dependent on the previous step, making the boundary of each historical gradient in SWA nested and the theoretical analysis even harder. To address the theoretical challenges, we adopt mathematical induction to find a recursive representation that bounds the gradient at each step. Based on this, we establish stability bounds supporting sampling with and without replacement in the non-convex setting. Furthermore, the derived generalization bounds of SWA are sharper than SGD. At last, experimental results on several benchmarks verify our theoretical results.

Data-Free Meta-Learning (DFML) aims to derive knowledge from a collection of pre-trained models without accessing their original data, enabling the rapid adaptation to new unseen tasks. Current methods often overlook the heterogeneity among pre-trained models, which leads to performance degradation due to task conflicts. In this paper, we empirically and theoretically identify and analyze the model heterogeneity in DFML. We find that model heterogeneity introduces a heterogeneity-homogeneity trade-off, where homogeneous models reduce task conflicts but also increase the overfitting risk. Balancing this trade-off is crucial for learning shared representations across tasks. Based on our findings, we propose Task Groupings Regularization, a novel approach that benefits from model heterogeneity by grouping and aligning conflicting tasks. Specifically, we embed pre-trained models into a task space to compute dissimilarity, and group heterogeneous models together based on this measure. Then, we introduce implicit gradient regularization within each group to mitigate potential conflicts. By encouraging a gradient direction suitable for all tasks, the meta-model captures shared representations that generalize across tasks. Comprehensive experiments showcase the superiority of our approach in multiple benchmarks, effectively tackling the model heterogeneity in challenging multi-domain and multi-architecture scenarios.

Multi-task learning (MTL) compresses the information from multiple tasks into a unified backbone to improve computational efficiency and generalization. Recent work directly merges multiple independently trained models to perform MTL instead of collecting their raw data for joint training, greatly expanding the application scenarios of MTL. However, by visualizing the representation distribution of existing model merging schemes, we find that the merged model often suffers from the dilemma of representation bias. That is, there is a significant discrepancy in the representation distribution between the merged and individual models, resulting in poor performance of merged MTL. In this paper, we propose a representation surgery solution called ``Surgery" to reduce representation bias in the merged model. Specifically, Surgery is a lightweight task-specific plugin that takes the representation of the merged model as input and attempts to output the biases contained in the representation from the merged model. We then designed an unsupervised optimization objective that updates the Surgery plugin by minimizing the distance between the merged model's representation and the individual model's representation. Extensive experiments demonstrate significant MTL performance improvements when our Surgery plugin is applied to state-of-the-art (SOTA) model merging schemes.

The recent advancements in cross-modal transformers have demonstrated their superior performance in RGB-D segmentation tasks by effectively integrating information from both RGB and depth modalities. However, existing methods often overlook the varying levels of informative content present in each modality, treating them equally and using models of the same architecture. This oversight can potentially hinder segmentation performance, especially considering that RGB images typically contain significantly more information than depth images. To address this issue, we propose PrimKD, a knowledge distillation based approach that focuses on guided multimodal fusion, with an emphasis on leveraging the primary RGB modality. In our approach, we utilize a model trained exclusively on the RGB modality as the teacher, guiding the learning process of a student model that fuses both RGB and depth modalities. To prioritize information from the primary RGB modality while leveraging the depth modality, we incorporate primary focused feature reconstruction and a selective alignment scheme. This integration enhances the overall freature fusion, resulting in improved segmentation results. We evaluate our proposed method on the NYU Depth V2 and SUN-RGBD datasets, and the experimental results demonstrate the effectiveness of PrimKD. Specifically, our approach achieves mIoU scores of 57.8 and 52.5 on these two datasets, respectively, surpassing existing counterparts by 1.5 and 0.4 mIoU. The code is available at https://github.com/xiaoshideta/PrimKD.

Transformer-based trajectory optimization methods have demonstrated exceptional performance in offline Reinforcement Learning (offline RL). Yet, it poses challenges due to substantial parameter size and limited scalability, which is particularly critical in sequential decision-making scenarios where resources are constrained such as in robots and drones with limited computational power. Mamba, a promising new linear-time sequence model, offers performance on par with transformers while delivering substantially fewer parameters on long sequences. As it remains unclear whether Mamba is compatible with trajectory optimization, this work aims to conduct comprehensive experiments to explore the potential of Decision Mamba (dubbed DeMa) in offline RL from the aspect of data structures and essential components with the following insights: (1) Long sequences impose a significant computational burden without contributing to performance improvements since DeMa's focus on sequences diminishes approximately exponentially. Consequently, we introduce a Transformer-like DeMa as opposed to an RNN-like DeMa. (2) For the components of DeMa, we identify the hidden attention mechanism as a critical factor in its success, which can also work well with other residual structures and does not require position embedding. Extensive evaluations demonstrate that our specially designed DeMa is compatible with trajectory optimization and surpasses previous methods, outperforming Decision Transformer (DT) with higher performance while using 30\% fewer parameters in Atari, and exceeding DT with only a quarter of the parameters in MuJoCo.

Backdoor attacks pose a significant threat to Deep Neural Networks (DNNs) as they allow attackers to manipulate model predictions with backdoor triggers. To address these security vulnerabilities, various backdoor purification methods have been proposed to purify compromised models. Typically, these purified models exhibit low Attack Success Rates (ASR), rendering them resistant to backdoored inputs. However, \textit{Does achieving a low ASR through current safety purification methods truly eliminate learned backdoor features from the pretraining phase?} In this paper, we provide an affirmative answer to this question by thoroughly investigating the \textit{Post-Purification Robustness} of current backdoor purification methods. We find that current safety purification methods are vulnerable to the rapid re-learning of backdoor behavior, even when further fine-tuning of purified models is performed using a very small number of poisoned samples. Based on this, we further propose the practical Query-based Reactivation Attack (QRA) which could effectively reactivate the backdoor by merely querying purified models. We find the failure to achieve satisfactory post-purification robustness stems from the insufficient deviation of purified models from the backdoored model along the backdoor-connected path. To improve the post-purification robustness, we propose a straightforward tuning defense, Path-Aware Minimization (PAM), which promotes deviation along backdoor-connected paths with extra model updates. Extensive experiments demonstrate that PAM significantly improves post-purification robustness while maintaining a good clean accuracy and low ASR. Our work provides a new perspective on understanding the effectiveness of backdoor safety tuning and highlights the importance of faithfully assessing the model's safety.

As a popular paradigm for juggling data privacy and collaborative training, federated learning (FL) is flourishing to distributively process the large scale of heterogeneous datasets on edged clients. Due to bandwidth limitations and security considerations, it ingeniously splits the original problem into multiple subproblems to be solved in parallel, which empowers primal dual solutions to great application values in FL. In this paper, we review the recent development of classical federated primal dual methods and point out a serious common defect of such methods in non-convex scenarios, which we say is a ``dual drift'' caused by dual hysteresis of those longstanding inactive clients under partial participation training. To further address this problem, we propose a novel Aligned Federated Primal Dual (A-FedPD) method, which constructs virtual dual updates to align global consensus and local dual variables for those protracted unparticipated local clients. Meanwhile, we provide a comprehensive analysis of the optimization and generalization efficiency for the A-FedPD method on smooth non-convex objectives, which confirms its high efficiency and practicality. Extensive experiments are conducted on several classical FL setups to validate the effectiveness of our proposed method.

Privacy protection of users' entire contribution of samples is important in distributed systems. The most effective approach is the two-stage scheme, which finds a small interval first and then gets a refined estimate by clipping samples into the interval. However, the clipping operation induces bias, which is serious if the sample distribution is heavy-tailed. Besides, users with large local sample sizes can make the sensitivity much larger, thus the method is not suitable for imbalanced users. Motivated by these challenges, we propose a Huber loss minimization approach to mean estimation under user-level differential privacy. The connecting points of Huber loss can be adaptively adjusted to deal with imbalanced users. Moreover, it avoids the clipping operation, thus significantly reducing the bias compared with the two-stage approach. We provide a theoretical analysis of our approach, which gives the noise strength needed for privacy protection, as well as the bound of mean squared error. The result shows that the new method is much less sensitive to the imbalance of user-wise sample sizes and the tail of sample distributions. Finally, we perform numerical experiments to validate our theoretical analysis.

Multi-task offline reinforcement learning aims to develop a unified policy for diverse tasks without requiring real-time interaction with the environment. Recent work explores sequence modeling, leveraging the scalability of the transformer architecture as a foundation for multi-task learning. Given the variations in task content and complexity, formulating policies becomes a challenging endeavor, requiring careful parameter sharing and adept management of conflicting gradients to extract rich cross-task knowledge from multiple tasks and transfer it to unseen tasks. In this paper, we propose the Decomposed Prompt Decision Transformer (DPDT) that adopts a two-stage paradigm to efficiently learn prompts for unseen tasks in a parameter-efficient manner. We incorporate parameters from pre-trained language models (PLMs) to initialize DPDT, thereby providing rich prior knowledge encoded in language models. During the decomposed prompt tuning phase, we learn both cross-task and task-specific prompts on training tasks to achieve prompt decomposition. In the test time adaptation phase, the cross-task prompt, serving as a good initialization, were further optimized on unseen tasks through test time adaptation, enhancing the model's performance on these tasks. Empirical evaluation on a series of Meta-RL benchmarks demonstrates the superiority of our approach. The project is available at https://github.com/ruthless-man/DPDT.

Transformers, originally devised for natural language processing (NLP), have also produced significant successes in computer vision (CV). Due to their strong expression power, researchers are investigating ways to deploy transformers for reinforcement learning (RL), and transformer-based models have manifested their potential in representative RL benchmarks. In this paper, we collect and dissect recent advances concerning the transformation of RL with transformers (transformer-based RL (TRL)) to explore the development trajectory and future trends of this field. We group the existing developments into two categories: architecture enhancements and trajectory optimizations, and examine the main applications of TRL in robotic manipulation, text-based games (TBGs), navigation, and autonomous driving. Architecture enhancement methods consider how to apply the powerful transformer structure to RL problems under the traditional RL framework, facilitating more precise modeling of agents and environments compared to traditional deep RL techniques. However, these methods are still limited by the inherent defects of traditional RL algorithms, such as bootstrapping and the "deadly triad". Trajectory optimization methods treat RL problems as sequence modeling problems and train a joint state-action model over entire trajectories under the behavior cloning framework; such approaches are able to extract policies from static datasets and fully use the long-sequence modeling capabilities of transformers. Given these advancements, the limitations and challenges in TRL are reviewed and proposals regarding future research directions are discussed. We hope that this survey can provide a detailed introduction to TRL and motivate future research in this rapidly developing field.

Continual learning (CL) aims to learn new tasks without forgetting previous tasks. However, existing CL methods require a large amount of raw data, which is often unavailable due to copyright considerations and privacy risks. Instead, stakeholders usually release pre-trained machine learning models as a service (MLaaS), which users can access via APIs. This paper considers two practical-yet-novel CL settings: data-efficient CL (DECL-APIs) and data-free CL (DFCL-APIs), which achieve CL from a stream of APIs with partial or no raw data. Performing CL under these two new settings faces several challenges: unavailable full raw data, unknown model parameters, heterogeneous models of arbitrary architecture and scale, and catastrophic forgetting of previous APIs. To overcome these issues, we propose a novel data-free cooperative continual distillation learning framework that distills knowledge from a stream of APIs into a CL model by generating pseudo data, just by querying APIs. Specifically, our framework includes two cooperative generators and one CL model, forming their training as an adversarial game. We first use the CL model and the current API as fixed discriminators to train generators via a derivative-free method. Generators adversarially generate hard and diverse synthetic data to maximize the response gap between the CL model and the API. Next, we train the CL model by minimizing the gap between the responses of the CL model and the black-box API on synthetic data, to transfer the API's knowledge to the CL model. Furthermore, we propose a new regularization term based on network similarity to prevent catastrophic forgetting of previous APIs. Our method performs comparably to classic CL with full raw data on the MNIST and SVHN datasets in the DFCL-APIs setting. In the DECL-APIs setting, our method achieves 0.97×, 0.75× and 0.69× performance of classic CL on the more challenging CIFAR10, CIFAR100, and MiniImageNet, respectively.

This paper delves into the problem of correlated time-series forecasting in practical applications, an area of growing interest in a multitude of fields such as stock price prediction and traffic demand analysis. Current methodologies primarily represent data using conventional graph structures, yet these fail to capture intricate structures with non-pairwise relationships. To address this challenge, we adopt dynamic hypergraphs in this study to better illustrate complex interactions, and introduce a novel hypergraph neural network model named CHNN for correlated time series forecasting. In more detail, CHNN leverages both semantic and topological similarities via an interaction model and hypergraph diffusion process, thereby constructing comprehensive collaborative correlation scores that effectively guide spatial message propagation. In addition, it incorporates short-term temporal information to generate efficient spatio-temporal feature maps. Lastly, a long-term temporal module is proposed to generate future predictions utilizing both temporal attention and a gated recurrent network. Comprehensive experiments conducted on four real-world datasets, i.e., Tiingo, Stocktwits, NYC-Taxi, and Social Network demonstrate that the proposed CHNN markedly outperforms a range of benchmark methods.

Lung cancer is the leading cause of cancer death worldwide. The best solution for lung cancer is to diagnose the pulmonary nodules in the early stage, which is usually accomplished with the aid of thoracic computed tomography (CT). As deep learning thrives, convolutional neural networks (CNNs) have been introduced into pulmonary nodule detection to help doctors in this labor-intensive task and demonstrated to be very effective. However, the current pulmonary nodule detection methods are usually domain-specific, and cannot satisfy the requirement of working in diverse real-world scenarios. To address this issue, we propose a slice grouped domain attention (SGDA) module to enhance the generalization capability of the pulmonary nodule detection networks. This attention module works in the axial, coronal, and sagittal directions. In each direction, we divide the input feature into groups, and for each group, we utilize a universal adapter bank to capture the feature subspaces of the domains spanned by all pulmonary nodule datasets. Then the bank outputs are combined from the perspective of domain to modulate the input group. Extensive experiments demonstrate that SGDA enables substantially better multi-domain pulmonary nodule detection performance compared with the state-of-the-art multi-domain learning methods.

Although few-shot learning aims to address data scarcity, it still requires large, annotated datasets for training, which are often unavailable due to cost and privacy concerns. Previous studies have utilized pre-trained diffusion models, either to synthesize auxiliary data besides limited labeled samples, or to employ diffusion models as zero-shot classifiers. However, they are limited to conditional diffusion models needing class prior information (e.g., carefully crafted text prompts) about unseen tasks. To overcome this, we leverage unconditional diffusion models without needs for class information to train a meta-model capable of generalizing to unseen tasks. The framework contains(1)a meta-learning without data approach that uses synthetic data during training; and(2)a diffusion model-based data augmentation to calibrate the distribution shift during testing. During meta-training, we implement aself-taughtclass-learner to gradually capture class concepts, guiding unconditional diffusion models to generate alabeledpseudo dataset. This pseudo dataset is then used to jointly train the class-learner and the meta-model, allowing for iterative refinement and clear differentiation between classes. During meta-testing, we introduce a data augmentation that employs the diffusion models used in meta-training, to narrow the gap between meta-training and meta-testing task distribution. This enables the meta-model trained onsyntheticimages to effectively classifyrealimages in unseen tasks. Comprehensive experiments showcase the superiority and adaptability of our approach in four real-world scenarios. Code available athttps://github.com/WalkerWorldPeace/MLWDUDM.

Multi-scenario and multi-task recommendation can use various feedback behaviors of users in different scenarios to learn users’ preferences and then make recommendations, which has attracted attention. However, the existing work ignores feature interactions and the fact that a pair of feature interactions will have differing levels of importance under different scenario-task pairs, leading to sub-optimal user preference learning. In this article, we propose a M ulti-scenario and M ulti-task aware F eature I nteraction model, dubbed MMFI , to explicitly model feature interactions and learn the importance of feature interaction pairs in different scenarios and tasks. Specifically, MMFI first incorporates a pairwise feature interaction unit and a scenario-task interaction unit to effectively capture the interaction of feature pairs and scenario-task pairs. Then MMFI designs a scenario-task aware attention layer for learning the importance of feature interactions from coarse-grained to fine-grained, improving the model’s performance on various scenario-task pairs. More specifically, this attention layer consists of three modules: a fully shared bottom module, a partially shared middle module, and a specific output module. Finally, MMFI adapts two sparsity-aware functions to remove some useless feature interactions. Extensive experiments on two public datasets demonstrate the superiority of the proposed method over the existing multi-task recommendation, multi-scenario recommendation, and multi-scenario & multi-task recommendation models.

Graph neural networks (GNNs) have achieved great success recently on graph classification tasks using supervised end-to-end training. Unfortunately, extensive noisy graph labels could exist in the real world because of the complicated processes of manual graph data annotations, which may significantly degrade the performance of GNNs. Therefore, we investigate the problem of graph classification with label noise, which is demanding because of the complex graph representation learning issue and serious memorization of noisy samples. In this work, we present a novel approach called S ubgra p h Set Netw or k with Sample Selection and Consis t ency Learning (SPORT) for this problem. To release the overfitting of GNNs, SPORT proposes to characterize each graph as a set of subgraphs generated by certain predefined stratagems, which can be viewed as samples from its underlying semantic distribution in graph space. Then we develop an equivariant network to encode the subgraph set with the consideration of the symmetry group. To further release the influences of noisy examples, we leverage the predictions of subgraphs to measure the likelihood of a sample being clean or noisy, followed by effective label updating. In addition, we propose a joint loss to advance the model generalizability by introducing consistency regularization. Comprehensive experiments on a wide range of graph classification datasets demonstrate the effectiveness of our SPORT. Specifically, SPORT outperforms the most competing baseline by up to 6.4%.

Federated learning (FL) can train collaboratively with several mobile terminals (MTs), which faces critical challenges in communication, resource, and privacy. Existing privacy-preserving methods usually adopt instance-level differential privacy (DP), which provides a rigorous privacy guarantee but with several bottlenecks: performance degradation, transmission overhead, and resource constraints. Therefore, we propose Fed-LTP, an efficient and privacy-enhanced FL framework withLotteryTicketHypothesis (LTH) and zero-concentrated DP(zCDP). It generates a pruned global model on the server side and conducts sparse-to-sparse training from scratch with zCDP on the client side. On the server side, two pruning schemes are proposed: (i) the weight-based pruning (LTH) determines the pruned global model structure; (ii) the iterative pruning further shrinks the size of the pruned model. Meanwhile, the performance of Fed-LTP is boosted via model validation based on the Laplace mechanism. On the client side, we use sparse-to-sparse training to solve the resource-constraints issue and provide tighter privacy analysis to reduce the privacy budget. We evaluate the effectiveness of Fed-LTP on several real-world datasets in both independent and identically distributed (IID) and non-IID settings. The results confirm the superiority of Fed-LTP over state-of-the-art (SOTA) methods in communication, computation, and memory efficiencies while realizing a better utility-privacy trade-off.

Despite remarkable successes in solving various complex decision-making tasks, training an imitation learning (IL) algorithm with deep neural networks (DNNs) suffers from the high-computational burden. In this work, we propose quantum IL (QIL) with a hope to utilize quantum advantage to speed up IL. Concretely, we develop two QIL algorithms: quantum behavioral cloning (Q-BC) and quantum generative adversarial IL (Q-GAIL). Q-BC is trained with a negative log-likelihood (NLL) loss in an offline manner that suits extensive expert data cases, whereas Q-GAIL works in an inverse reinforcement learning (IRL) scheme, which is online, on-policy, and is suitable for limited expert data cases. For both QIL algorithms, we adopt variational quantum circuits (VQCs) in place of DNNs for representing policies, which are modified with data reuploading and scaling parameters to enhance the expressivity. We first encode classical data into quantum states as inputs, then perform VQCs, and finally measure quantum outputs to obtain control signals of agents. Experiment results demonstrate that both Q-BC and Q-GAIL can achieve comparable performance compared to classical counterparts, with the potential of quantum speedup. To our knowledge, we are the first to propose the concept of QIL and conduct pilot studies, which paves the way for the quantum era.

Federated learning (FL) is a promising framework for privacy-preserving and distributed training with decentralized clients. However, there exists a large divergence between the collected local updates and the expected global update, which is known as the client drift and mainly caused by heterogeneous data distribution among clients, multiple local training steps, and partial client participation training. Most existing works tackle this challenge based on the empirical risk minimization (ERM) rule, while less attention has been paid to the relationship between the global loss landscape and the generalization ability. In this work, we propose FedGAMMA, a novel FL algorithm with Global sharpness-Aware MiniMizAtion to seek a global flat landscape with high performance. Specifically, in contrast to FedSAM which only seeks the local flatness and still suffers from performance degradation when facing the client-drift issue, we adopt a local varieties control technique to better align each client's local updates to alleviate the client drift and make each client heading toward the global flatness together. Finally, extensive experiments demonstrate that FedGAMMA can substantially outperform several existing FL baselines on various datasets, and it can well address the client-drift issue and simultaneously seek a smoother and flatter global landscape.

We propose a novel master-slave architecture to solve the top- combinatorial multiarmed bandits (CMABs) problem with nonlinear bandit feedback and diversity constraints, which, to the best of our knowledge, is the first combinatorial bandits setting considering diversity constraints under bandit feedback. Specifically, to efficiently explore the combinatorial and constrained action space, we introduce six slave models with distinguished merits to generate diversified samples well balancing rewards and constraints as well as efficiency. Moreover, we propose teacher learning-based optimization and the policy cotraining technique to boost the performance of the multiple slave models. The master model then collects the elite samples provided by the slave models and selects the best sample estimated by a neural contextual UCB-based network (NeuralUCB) to decide on a tradeoff between exploration and exploitation. Thanks to the elaborate design of slave models, the cotraining mechanism among slave models, and the novel interactions between the master and slave models, our approach significantly surpasses existing state-of-the-art algorithms in both synthetic and real datasets for recommendation tasks. The code is available at https://github.com/huanghanchi/Master-slave-Algorithm-for-Top-K-Bandits.

Integrating adaptive learning rate and momentum techniques into stochastic gradient descent (SGD) leads to a large class of efficiently accelerated adaptive stochastic algorithms, such as AdaGrad, RMSProp, Adam, AccAdaGrad, and so on. In spite of their effectiveness in practice, there is still a large gap in their theories of convergences, especially in the difficult nonconvex stochastic setting. To fill this gap, we propose weighted AdaGrad with unified momentum and dubbed AdaUSM, which has the main characteristics that: 1) it incorporates a unified momentum scheme that covers both the heavy ball (HB) momentum and the Nesterov accelerated gradient (NAG) momentum and 2) it adopts a novel weighted adaptive learning rate that can unify the learning rates of AdaGrad, AccAdaGrad, Adam, and RMSProp. Moreover, when we take polynomially growing weights in AdaUSM, we obtain its O(log(T)/√T) convergence rate in the nonconvex stochastic setting. We also show that the adaptive learning rates of Adam and RMSProp correspond to taking exponentially growing weights in AdaUSM, thereby providing a new perspective for understanding Adam and RMSProp. Finally, comparative experiments of AdaUSM against SGD with momentum, AdaGrad, AdaEMA, Adam, and AMSGrad on various deep learning models and datasets are also carried out.

The field of multimedia research has witnessed significant interest in leveraging multimodal pretrained neural network models to perceive and represent the physical world. Among these models, vision-language pretraining (VLP) has emerged as a captivating topic. Currently, the prevalent approach in VLP involves supervising the training process with paired image-text data. However, limited efforts have been dedicated to exploring the extraction of essential linguistic knowledge, such as semantics and syntax, during VLP and understanding its impact on multimodal alignment. In response, our study aims to shed light on the influence of comprehensive linguistic knowledge encompassing semantic expression and syntactic structure on multimodal alignment. To achieve this, we introduce SNARE , a large-scale multimodal alignment probing benchmark designed specifically for the detection of vital linguistic components, including lexical, semantic, and syntax knowledge. SNARE offers four distinct tasks: Semantic Structure, Negation Logic, Attribute Ownership, and Relationship Composition. Leveraging SNARE , we conduct holistic analyses of six advanced VLP models (BLIP, CLIP, Flava, X-VLM, BLIP2, and GPT-4), along with human performance, revealing key characteristics of the VLP model: (i) Insensitivity to complex syntax structures, relying primarily on content words for sentence comprehension. (ii) Limited comprehension of sentence combinations and negations. (iii) Challenges in determining actions or spatial relations within visual information, as well as difficulties in verifying the correctness of ternary relationships. Based on these findings, we propose the following strategies to enhance multimodal alignment in VLP: (1) Utilize a large generative language model as the language backbone in VLP to facilitate the understanding of complex sentences. (2) Establish high-quality datasets that emphasize content words and employ simple syntax, such as short-distance semantic composition, to improve multimodal alignment. (3) Incorporate more fine-grained visual knowledge, such as spatial relationships, into pretraining objectives. 1

Personalized federated learning (PFL) is a framework that targets individual models for optimization, providing better privacy and flexibility for clients. However, in challenging intelligent sensing applications, the heterogeneous client’s data distributions make the aggregation of local models in the server unstable or even hard to converge. To deal with the performance degradation caused by the preceding problem, existing PFL methods focus more on how to fine-tune the global model but ignore the impact of the global model fusion algorithm on the results. In this article, we propose a new explainable neural-aware decoupling fusion based PFL framework, p-FedADF , to address the preceding challenges. It contains two carefully designed modules. The local decoupling module, deployed on the client, utilizes the architecture disentangle technique to decouple the feature extractors in the client’s local model into sub-network according to data categories. It obtains the inference process of feature extraction for different categories of data by training. The global aggregation module, deployed on the server, aligns the sub-network positions for multiple clients and implements a fine-grained generic feature extractor aggregation. In addition, we provide a mask encoding scheme to reduce the communication overhead of transmitting the sub-network sets between the server and clients. Our p-FedADF obtains 1.6%, 0.2%, 2.3%, and 4.5% improvement on a real-world dataset and three benchmark datasets, compared to state-of-the-art methods.

As an essential resource management problem in network virtualization, virtual network embedding (VNE) aims to allocate the finite resources of physical network to sequentially arriving virtual network requests (VNRs) with different resource demands. Since this is an NP-hard combinatorial optimization problem, many efforts have been made to provide viable solutions. However, most existing approaches have either ignored the admission control of VNRs, which has a potential impact on long-term performances, or not fully exploited the temporal and topological features of the physical network and VNRs. In this paper, we propose a deepHierarchicalReinforcementLearning approach to learn a jointAdmissionControl andResourceAllocation policy for VNE, named HRL-ACRA. Specifically, the whole VNE process is decomposed into an upper-level policy for deciding whether to admit the arriving VNR or not and a lower-level policy for allocating resources of the physical network to meet the requirement of VNR through the HRL approach. Considering the proximal policy optimization as the basic training algorithm, we also adopt the average reward method to address the infinite horizon problem of the upper-level agent and design a customized multi-objective intrinsic reward to alleviate the sparse reward issue of the lower-level agent. Moreover, we develop a deep feature-aware graph neural network to capture the features of VNR and physical network and exploit a sequence-to-sequence model to generate embedding actions iteratively. Finally, extensive experiments are conducted in various settings, and show that HRL-ACRA outperforms state-of-the-art baselines in terms of both the acceptance ratio and long-term average revenue. Our code is available athttps://github.com/GeminiLight/hrl-acra.

Recently, to reap the quantum advantage, empowering reinforcement learning (RL) with quantum computing has attracted much attention, which is dubbed as quantum RL (QRL). However, current QRL algorithms employ an online learning scheme, i.e., the policy that is run on a quantum computer needs to interact with the environment to collect experiences, which could be expensive and dangerous for practical applications. In this paper, we aim to solve this problem in an offline learning manner. To be more specific, we develop the first offline quantum RL (offline QRL) algorithm named CQ2L (Conservative Quantum Q-learning), which learns from offline samples and does not require any interaction with the environment. CQ2L utilizes variational quantum circuits (VQCs), which are improved with data re-uploading and scaling parameters, to represent Q-value functions of agents. To suppress the overestimation of Q-values resulting from offline data, we first employ a double Q-learning framework to reduce the overestimation bias; then a penalty term that encourages generating conservative Q-values is designed. We conduct abundant experiments to demonstrate that the proposed method CQ2L can successfully solve offline QRL tasks that the online counterpart could not.

Federated learning aims to collaboratively train models without accessing their client's local private data. The data may be Non-IID for different clients and thus resulting in poor performance. Recently, personalized federated learning (PFL) has achieved great success in handling Non-IID data by enforcing regularization in local optimization or improving the model aggregation scheme on the server. However, most of the PFL approaches do not take into account the unfair competition issue caused by the imbalanced data distribution and lack of positive samples for some classes in each client. To address this issue, we propose a novel and generic PFL framework termed Federated Averaging via Binary Classification, dubbed FedABC. In particular, we adopt the ``one-vs-all'' training strategy in each client to alleviate the unfair competition between classes by constructing a personalized binary classification problem for each class. This may aggravate the class imbalance challenge and thus a novel personalized binary classification loss that incorporates both the under-sampling and hard sample mining strategies is designed. Extensive experiments are conducted on two popular datasets under different settings, and the results demonstrate that our FedABC can significantly outperform the existing counterparts.

Multi-task learning (MTL) models have demonstrated impressive results in computer vision, natural language processing, and recommender systems. Even though many approaches have been proposed, how well these approaches balance different tasks on each parameter still remains unclear. In this paper, we propose to measure the task dominance degree of a parameter by the total updates of each task on this parameter. Specifically, we compute the total updates by the exponentially decaying Average of the squared Updates (AU) on a parameter from the corresponding task. Based on this novel metric, we observe that many parameters in existing MTL methods, especially those in the higher shared layers, are still dominated by one or several tasks. The dominance of AU is mainly due to the dominance of accumulative gradients from one or several tasks. Motivated by this, we propose a Task-wise Adaptive learning rate approach, AdaTask in short, to separate the accumulative gradients and hence the learning rate of each task for each parameter in adaptive learning rate approaches (e.g., AdaGrad, RMSProp, and Adam). Comprehensive experiments on computer vision and recommender system MTL datasets demonstrate that AdaTask significantly improves the performance of dominated tasks, resulting SOTA average task-wise performance. Analysis on both synthetic and real-world datasets shows AdaTask balance parameters in every shared layer well.

Safety comes first in many real-world applications involving autonomous agents. Despite a large number of reinforcement learning (RL) methods focusing on safety-critical tasks, there is still a lack of high-quality evaluation of those algorithms that adheres to safety constraints at each decision step under complex and unknown dynamics. In this paper, we revisit prior work in this scope from the perspective of state-wise safe RL and categorize them as projection-based, recovery-based, and optimization-based approaches, respectively. Furthermore, we propose Unrolling Safety Layer (USL), a joint method that combines safety optimization and safety projection. This novel technique explicitly enforces hard constraints via the deep unrolling architecture and enjoys structural advantages in navigating the trade-off between reward improvement and constraint satisfaction. To facilitate further research in this area, we reproduce related algorithms in a unified pipeline and incorporate them into SafeRL-Kit, a toolkit that provides off-the-shelf interfaces and evaluation utilities for safety-critical tasks. We then perform a comparative study of the involved algorithms on six benchmarks ranging from robotic control to autonomous driving. The empirical results provide an insight into their applicability and robustness in learning zero-cost-return policies without task-dependent handcrafting. The project page is available at https://sites.google.com/view/saferlkit.

The goal of data-free meta-learning is to learn useful prior knowledge from a collection of pre-trained models without accessing their training data. However, existing works only solve the problem in parameter space, which (i) ignore the fruitful data knowledge contained in the pretrained models; (ii) can not scale to large-scale pre-trained models; (iii) can only meta-learn pre-trained models with the same network architecture. To address those issues, we propose a unified framework, dubbed PURER, which contains: (1) ePisode cUrriculum inveRsion (ECI) during data-free meta training; and (2) invErsion calibRation following inner loop (ICFIL) during meta testing. During meta training, we propose ECI to perform pseudo episode training for learning to adapt fast to new unseen tasks. Specifically, we progressively synthesize a sequence of pseudo episodes by distilling the training data from each pre-trained model. The ECI adaptively increases the difficulty level of pseudo episodes according to the real-time feedback of the meta model. We formulate the optimization process of meta training with ECI as an adversarial form in an end-to-end manner. During meta testing, we further propose a simple plug-and-play supplement—ICFIL—only used during meta testing to narrow the gap between meta training and meta testing task distribution. Extensive experiments in various real-world scenarios show the superior performance of ours.

Robust generalization aims to tackle the most challenging data distributions which are rare in the training set and contain severe noises, i.e., photon-limited corruptions. Common solutions such as distributionally robust optimization (DRO) focus on the worst-case empirical risk to ensure low training error on the uncommon noisy distributions. However, due to the over-parameterized model being optimized on scarce worst-case data, DRO fails to produce a smooth loss landscape, thus struggling on generalizing well to the test set. Therefore, instead of focusing on the worst-case risk minimization, we propose SharpDRO by penalizing the sharpness of the worst-case distribution, which measures the loss changes around the neighbor of learning parameters. Through worst-case sharpness minimization, the proposed method successfully produces a flat loss curve on the corrupted distributions, thus achieving robust generalization. Moreover, by considering whether the distribution annotation is available, we apply SharpDRO to two problem settings and design a worst-case selection process for robust generalization. Theoretically, we show that SharpDRO has a great convergence guarantee. Experimentally, we simulate photon-limited corruptions using CIFAR10/100 and ImageNet30 datasets and show that SharpDRO exhibits a strong generalization ability against severe corruptions and exceeds well-known baseline methods with large performance gains.

To defend the inference attacks and mitigate the sensitive information leakages in Federated Learning (FL), clientlevel Differentially Private FL (DPFL) is the de-facto standard for privacy protection by clipping local updates and adding random noise. However, existing DPFL methods tend to make a sharper loss landscape and have poorer weight perturbation robustness, resulting in severe performance degradation. To alleviate these issues, we propose a novel DPFL algorithm named DP-FedSAM, which leverages gradient perturbation to mitigate the negative impact of DP. Specifically, DP-FedSAM integrates Sharpness Aware Minimization (SAM) optimizer to generate local flatness models with better stability and weight perturbation robustness, which results in the small norm of local updates and robustness to DP noise, thereby improving the performance. From the theoretical perspective, we analyze in detail how DP-FedSAM mitigates the performance degradation induced by DP. Meanwhile, we give rigorous privacy guarantees with Rényi DP and present the sensitivity analysis of local updates. At last, we empirically confirm that our algorithm achieves state-of-the-art (SOTA) performance compared with existing SOTA baselines in DPFL.

Continual Learning (CL) has achieved rapid progress in recent years. However, it is still largely unknown how to determine whether a CL model is trustworthy and how to foster its trustworthiness. This work focuses on evaluating and improving the robustness to corruptions of existing CL models. Our empirical evaluation results show that existing state-of-the-art (SOTA) CL models are particularly vulnerable to various data corruptions during testing. To make them trustworthy and robust to corruptions deployed in safety-critical scenarios, we propose a meta-learning framework of self-adaptive data augmentation to tackle the corruption robustness in CL. The proposed framework, MetaMix, learns to augment and mix data, automatically transforming the new task data or memory data. It directly optimizes the generalization performance against data corruptions during training. To evaluate the corruption robustness of our proposed approach, we construct several CL corruption datasets with different levels of severity. We perform comprehensive experiments on both task- and class-continual learning. Extensive experiments demonstrate the effectiveness of our proposed method compared to SOTA baselines.

Scaling the size of language models usually leads to remarkable advancements in NLP tasks.But it often comes with a price of growing computational cost.Although a sparse Mixture of Experts (MoE) can reduce the cost by activating a small subset of parameters (e.g., one expert) for each input, its computation escalates significantly if increasing the number of activated experts, limiting its practical utility.Can we retain the advantages of adding more experts without substantially increasing the computational costs?In this paper, we first demonstrate the superiority of selecting multiple experts and then propose a computation-efficient approach called Merging Experts into One (MEO), which reduces the computation cost to that of a single expert.Extensive experiments show that MEO significantly improves computational efficiency, e.g., FLOPS drops from 72.0G of vanilla MoE to 28.9G (MEO).Moreover, we propose a token-level attention block that further enhances the efficiency and performance of token-level MEO, e.g., 83.3% (MEO) vs. 82.6%(vanilla MoE) average score on the GLUE benchmark.Our code will be released upon acceptance.Code will be released at

Quantization is a promising approach for reducing memory overhead and accelerating inference, especially in large pre-trained language model (PLM) scenarios.While having no access to original training data due to security and privacy concerns has emerged the demand for zero-shot quantization.Most of the cuttingedge zero-shot quantization methods primarily ❶ apply to computer vision tasks, and ❷ neglect of overfitting problem in the generative adversarial learning process, leading to sub-optimal performance.Motivated by this, we propose a novel zero-shot sharpness-aware quantization (ZSAQ) framework for the zeroshot quantization of various PLMs.The key algorithm in solving ZSAQ is the SAM-SGA optimization, which aims to improve the quantization accuracy and model generalization via optimizing a minimax problem.We theoretically prove the convergence rate for the minimax optimization problem and this result can be applied to other nonconvex-PL minimax optimization frameworks.Extensive experiments on 11 tasks demonstrate that our method brings consistent and significant performance gains on both discriminative and generative PLMs, i.e., up to +6.98 average score.Furthermore, we empirically validate that our method can effectively improve the model generalization.

The goal of continual learning (CL) is to continuously learn new tasks without forgetting previously learned old tasks. To alleviate catastrophic forgetting, gradient projection based CL methods require that the gradient updates of new tasks are orthogonal to the subspace spanned by old tasks. This limits the learning process and leads to poor performance on the new task due to the projection constraint being too strong. In this paper, we first revisit the gradient projection method from the perspective of flatness of loss surface, and find that unflatness of the loss surface leads to catastrophic forgetting of the old tasks when the projection constraint is reduced to improve the performance of new tasks. Based on our findings, we propose a Data Augmented Flatness-aware Gradient Projection (DFGP) method to solve the problem, which consists of three modules: data and weight perturbation, flatness-aware optimization, and gradient projection. Specifically, we first perform a flatness-aware perturbation on the task data and current weights to find the case that makes the task loss worst. Next, flatness-aware optimization optimizes both the loss and the flatness of the loss surface on raw and worst-case perturbed data to obtain a flatness-aware gradient. Finally, gradient projection updates the network with the flatness-aware gradient along directions orthogonal to the subspace of the old tasks. Extensive experiments on four datasets show that our method improves the flatness of loss surface and the performance of new tasks, and achieves state-of-the-art (SOTA) performance in the average accuracy of all tasks.

Federated learning (FL) is a prevalent distributed machine learning approach that enables collaborative training of a global model across multiple devices without sharing local data. However, the presence of long-tailed data can negatively deteriorate the model’s performance in real-world FL applications. Moreover, existing re-balance strategies are less effective for the federated long-tailed issue when directly utilizing local label distribution as the class prior at the clients’ side. To this end, we propose a novel Global Balanced Multi-Expert (GBME) framework to optimize a balanced global objective, which does not require additional information beyond the standard FL pipeline. In particular, a proxy is derived from the accumulated gradients uploaded by the clients after local training, and is shared by all clients as the class prior for re-balance training. Such a proxy can also guide the client grouping to train a multi-expert model, where the knowledge from different clients can be aggregated via the ensemble of different experts corresponding to different client groups. To further strengthen the privacy-preserving ability, we present a GBME-p algorithm with a theoretical guarantee to prevent privacy leakage from the proxy. Extensive experiments on long-tailed decentralized datasets demonstrate the effectiveness of GBME and GBME-p, both of which show superior performance to state-of-the-art methods. The code is available at here.

Backdoor defense, which aims to detect or mitigate the effect of malicious triggers introduced by attackers, is becoming increasingly critical for machine learning security and integrity. Fine-tuning based on benign data is a natural defense to erase the backdoor effect in a backdoored model. However, recent studies show that, given limited benign data, vanilla fine-tuning has poor defense performance. In this work, we firstly investigate the vanilla fine-tuning process for backdoor mitigation from the neuron weight perspective, and find that backdoor-related neurons are only slightly perturbed in the vanilla fine-tuning process, which explains its poor backdoor defense performance. To enhance the fine-tuning based defense, inspired by the observation that the backdoor-related neurons often have larger weight norms, we propose FT-SAM, a novel backdoor defense paradigm that aims to shrink the norms of backdoor-related neurons by incorporating sharpness-aware minimization with fine-tuning. We demonstrate the effectiveness of our method on several benchmark datasets and network architectures, where it achieves state-of-the-art defense performance, and provide extensive analysis to reveal the FT-SAM’s mechanism. Overall, our work provides a promising avenue for improving the robustness of machine learning models against backdoor attacks. Codes are available at https://github.com/SCLBD/BackdoorBench.

Data-free meta-learning (DFML) aims to enable efficient learning of new tasks by meta-learning from a collection of pre-trained models without access to the training data. Existing DFML work can only meta-learn from (i) white-box and (ii) small-scale pre-trained models (iii) with the same architecture, neglecting the more practical setting where the users only have inference access to the APIs with arbitrary model architectures and model scale inside. To solve this issue, we propose a Bi-level Data-free Meta Knowledge Distillation (BiDf-MKD) framework to transfer more general meta knowledge from a collection of black-box APIs to one single meta model. Specifically, by just querying APIs, we inverse each API to recover its training data via a zero-order gradient estimator and then perform meta-learning via a novel bi-level meta knowledge distillation structure, in which we design a boundary query set recovery technique to recover a more informative query set near the decision boundary. In addition, to encourage better generalization within the setting of limited API budgets, we propose task memory replay to diversify the underlying task distribution by covering more interpolated tasks. Extensive experiments in various real-world scenarios show the superior performance of our BiDf-MKD framework.

This paper reveals a new appeal of the recently emerged large-kernel Convolutional Neural Networks (ConvNets): as the teacher in Knowledge Distillation (KD) for small-kernel ConvNets. While Transformers have led state-of-the-art (SOTA) performance in various fields with ever-larger models and labeled data, small-kernel ConvNets are considered more suitable for resource-limited applications due to the efficient convolution operation and compact weight sharing. KD is widely used to boost the performance of small-kernel ConvNets. However, previous research shows that it is not quite effective to distill knowledge (e.g., global information) from Transformers to small-kernel ConvNets, presumably due to their disparate architectures. We hereby carry out a first-of-its-kind study unveiling that modern large-kernel ConvNets, a compelling competitor to Vision Transformers, are remarkably more effective teachers for small-kernel ConvNets, due to more similar architectures. Our findings are backed up by extensive experiments on both logit-level and feature-level KD "out of the box", with no dedicated architectural nor training recipe modifications. Notably, we obtain the **best-ever pure ConvNet** under 30M parameters with 83.1% top-1 accuracy on ImageNet, outperforming current SOTA methods including ConvNeXt V2 and Swin V2. We also find that beneficial characteristics of large-kernel ConvNets, e.g., larger effective receptive fields, can be seamlessly transferred to students through this large-to-small kernel distillation. Code is available at: https://github.com/VITA-Group/SLaK.

To mitigate the privacy leakages and communication burdens of Federated Learning (FL), decentralized FL (DFL) discards the central server and each client only communicates with its neighbors in a decentralized communication network. However, existing DFL suffers from high inconsistency among local clients, which results in severe distribution shift and inferior performance compared with centralized FL (CFL), especially on heterogeneous data or sparse communication topologies. To alleviate this issue, we propose two DFL algorithms named DFedSAM and DFedSAM-MGS to improve the performance of DFL. Specifically, DFedSAM leverages gradient perturbation to generate local flat models via Sharpness Aware Minimization (SAM), which searches for models with uniformly low loss values. DFedSAM-MGS further boosts DFedSAM by adopting Multiple Gossip Steps (MGS) for better model consistency, which accelerates the aggregation of local flat models and better balances communication complexity and generalization. Theoretically, we present improved convergence rates $\small \mathcal{O}\big(\frac{1}{\sqrt{KT}}+\frac{1}{T}+\frac{1}{K^{1/2}T^{3/2}(1-\lambda)^2}\big)$ and $\small \mathcal{O}\big(\frac{1}{\sqrt{KT}}+\frac{1}{T}+\frac{\lambda^Q+1}{K^{1/2}T^{3/2}(1-\lambda^Q)^2}\big)$ in non-convex setting for DFedSAM and DFedSAM-MGS, respectively, where $1-\lambda$ is the spectral gap of gossip matrix and $Q$ is the number of MGS. Empirically, our methods can achieve competitive performance compared with CFL methods and outperform existing DFL methods.

In federated learning (FL), a cluster of local clients are chaired under the coordination of the global server and cooperatively train one model with privacy protection. Due to the multiple local updates and the isolated non-iid dataset, clients are prone to overfit into their own optima, which extremely deviates from the global objective and significantly undermines the performance. Most previous works only focus on enhancing the consistency between the local and global objectives to alleviate this prejudicial client drifts from the perspective of the optimization view, whose performance would be prominently deteriorated on the high heterogeneity. In this work, we propose a novel and general algorithm FedSMOO by jointly considering the optimization and generalization targets to efficiently improve the performance in FL. Concretely, FedSMOO adopts a dynamic regularizer to guarantee the local optima towards the global objective, which is meanwhile revised by the global Sharpness Aware Minimization (SAM) optimizer to search for the consistent flat minima. Our theoretical analysis indicates that FedSMOO achieves fast $\mathcal{O}(1/T)$ convergence rate with low generalization bound. Extensive numerical studies are conducted on the real-world dataset to verify its peerless efficiency and excellent generality.

Although graph neural networks (GNNs) have achieved impressive achievements in graph classification, they often need abundant task-specific labels, which could be extensively costly to acquire. A credible solution is to explore additional labeled graphs to enhance unsupervised learning on the target domain. However, how to apply GNNs to domain adaptation remains unsolved owing to the insufficient exploration of graph topology and the significant domain discrepancy. In this paper, we propose Coupled Contrastive Graph Representation Learning (CoCo), which extracts the topological information from coupled learning branches and reduces the domain discrepancy with coupled contrastive learning. CoCo contains a graph convolutional network branch and a hierarchical graph kernel network branch, which explore graph topology in implicit and explicit manners. Besides, we incorporate coupled branches into a holistic multi-view contrastive learning framework, which not only incorporates graph representations learned from complementary views for enhanced understanding, but also encourages the similarity between cross-domain example pairs with the same semantics for domain alignment. Extensive experiments on popular datasets show that our CoCo outperforms these competing baselines in different settings generally.

Recently, various successful applications utilizing expert states in imitation learning (IL) have been witnessed. However, IL from visual inputs (ILfVI), which has a greater promise to be widely applied by using online visual resources, suffers from low data-efficiency and poor performance resulted from on-policy learning and high-dimensional visual inputs. We propose OPIfVI (Off-Policy Imitation from Visual Inputs), which is composed of an off-policy learning manner, data augmentation, and encoder techniques, to tackle the mentioned challenges, respectively. More specifically, to improve data-efficiency, OPIfVI conducts IL in an off-policy manner, with which sampled data used multiple times. In addition, we enhance the stability of OPIfVI with spectral normalization to mitigate the side effect of off-policy training. The core factor, contributing to the poor performance of ILfVI, that we think is agents could not extract meaningful features from visual inputs. Hence, OPIfVI employs data augmentation from computer vision to help train encoders to better extract features from visual inputs. Besides, a specific structure of gradient backpropagation for the encoder is designed to stabilize the encoder training. At last, we demonstrate that OPIfVI can achieve expert-level performance and outperform existing baselines via extensive experiments using DeepMind Control Suite.

Recently, the efficient deployment and acceleration of powerful vision transformers (ViTs) on resource-limited edge devices for providing multimedia services have become attractive tasks. Although early exiting is a feasible solution for accelerating inference, most works focus on convolutional neural networks (CNNs) and transformer models in natural language processing (NLP). Moreover, the direct application of early exiting methods to ViTs may result in substantial performance degradation. To tackle this challenge, we systematically investigate the efficacy of early exiting in ViTs and point out that the insufficient feature representations in shallow internal classifiers and the limited ability to capture target semantic information in deep internal classifiers restrict the performance of these methods. We then propose an early exiting framework for general ViTs termed LGViT, which incorporates heterogeneous exiting heads, namely, local perception head and global aggregation head, to achieve an efficiency-accuracy trade-off. In particular, we develop a novel two-stage training scheme, including end-to-end training and self-distillation with the backbone frozen to generate early exiting ViTs, which facilitates the fusion of global and local information extracted by the two types of heads. We conduct extensive experiments using three popular ViT backbones on three vision datasets. Results demonstrate that our LGViT can achieve competitive performance with approximately 1.8 × speed-up.

Sparse training has received an upsurging interest in machine learning due to its tantalizing saving potential for both the entire training process as well as the inference. Dynamic sparse training (DST) as a leading approach can train deep neural networks at high sparsity from scratch to match the performance of their dense counterparts. However, most if not all DST prior arts demonstrate their effectiveness on unstructured sparsity with highly irregular sparse patterns, which receives limited support in common hardware. This limitation hinders the usage of DST in practice. In this paper, we propose Channel-aware dynamic sparse (Chase), that for the first time seamlessly translates the promise of unstructured dynamic sparsity to GPU-friendly channel-level sparsity (not fine-grained N:M or group sparsity) during one end-to-end training process, without any ad-hoc operations. The resulting small sparse networks can be directly accelerated by commodity hardware, without using any particularly sparsity-aware hardware accelerators. This appealing outcome is partially motivated by a hidden phenomenon of dynamic sparsity: off-the-shelf unstructured DST implicitly involves biased parameter reallocation across channels, with a large fraction of channels (up to 60%) being sparser than others. By progressively identifying and removing these channels during training, our approach transfers unstructured sparsity to channel-wise sparsity. Our experimental results demonstrate that Chase achieves 1.7x inference throughput speedup on common GPU devices without compromising accuracy with ResNet-50 on ImageNet. We release our code in https://github.com/luuyin/chase.

Federated Learning (FL) is an emerging collaborative machine learning framework where multiple clients train the global model without sharing their own datasets. In FL, the model inconsistency caused by the local data heterogeneity across clients results in the near-orthogonality of client updates, which leads to the global update norm reduction and slows down the convergence. Most previous works focus on eliminating the difference of parameters (or gradients) between the local and global models, which may fail to reflect the model inconsistency due to the complex structure of the machine learning model and the Euclidean space's limitation in meaningful geometric representations. In this paper, we propose FedMRUR by adopting the manifold model fusion scheme and a new global optimizer to alleviate the negative impacts. Concretely, FedMRUR adopts a hyperbolic graph manifold regularizer enforcing the representations of the data in the local and global models are close to each other in a low-dimensional subspace. Because the machine learning model has the graph structure, the distance in hyperbolic space can reflect the model bias better than the Euclidean distance. In this way, FedMRUR exploits the manifold structures of the representations to significantly reduce the model inconsistency. FedMRUR also aggregates the client updates norms as the global update norm, which can appropriately enlarge each client's contribution to the global update, thereby mitigating the norm reduction introduced by the near-orthogonality of client updates. Furthermore, we theoretically prove that our algorithm can achieve a linear speedup property $\mathcal{O}(\frac{1}{\sqrt{SKT}})$ for non-convex setting under partial client participation, where $S$ is the participated clients number, $K$ is the local interval and $T$ is the total number of communication rounds. Experiments demonstrate that FedMRUR can achieve a new state-of-the-art (SOTA) accuracy with less communication.

Semi-Supervised Learning (SSL) has been an effective way to leverage abundant unlabeled data with extremely scarce labeled data. However, most SSL methods are commonly based on instance-wise consistency between different data transformations. Therefore, the label guidance on labeled data is hard to be propagated to unlabeled data. Consequently, the learning process on labeled data is much faster than on unlabeled data which is likely to fall into a local minima that does not favor unlabeled data, leading to sub-optimal generalization performance. In this paper, we propose FlatMatch which minimizes a cross-sharpness measure to ensure consistent learning performance between the two datasets. Specifically, we increase the empirical risk on labeled data to obtain a worst-case model which is a failure case needing to be enhanced. Then, by leveraging the richness of unlabeled data, we penalize the prediction difference (i.e., cross-sharpness) between the worst-case model and the original model so that the learning direction is beneficial to generalization on unlabeled data. Therefore, we can calibrate the learning process without being limited to insufficient label information. As a result, the mismatched learning performance can be mitigated, further enabling the effective exploitation of unlabeled data and improving SSL performance. Through comprehensive validation, we show FlatMatch achieves state-of-the-art results in many SSL settings.

Data augmentation (DA) is a crucial technique for enhancing the sample efficiency of visual reinforcement learning (RL) algorithms. Notably, employing simple observation transformations alone can yield outstanding performance without extra auxiliary representation tasks or pre-trained encoders. However, it remains unclear which attributes of DA account for its effectiveness in achieving sample-efficient visual RL. To investigate this issue and further explore the potential of DA, this work conducts comprehensive experiments to assess the impact of DA's attributes on its efficacy and provides the following insights and improvements: (1) For individual DA operations, we reveal that both ample spatial diversity and slight hardness are indispensable. Building on this finding, we introduce Random PadResize (Rand PR), a new DA operation that offers abundant spatial diversity with minimal hardness. (2) For multi-type DA fusion schemes, the increased DA hardness and unstable data distribution result in the current fusion schemes being unable to achieve higher sample efficiency than their corresponding individual operations. Taking the non-stationary nature of RL into account, we propose a RL-tailored multi-type DA fusion scheme called Cycling Augmentation (CycAug), which performs periodic cycles of different DA operations to increase type diversity while maintaining data distribution consistency. Extensive evaluations on the DeepMind Control suite and CARLA driving simulator demonstrate that our methods achieve superior sample efficiency compared with the prior state-of-the-art methods.

It has been widely observed that deep neural networks (DNN) are vulnerable to backdoor attacks where attackers could manipulate the model behavior maliciously by tampering with a small set of training samples. Although a line of defense methods is proposed to mitigate this threat, they either require complicated modifications to the training process or heavily rely on the specific model architecture, which makes them hard to deploy into real-world applications. Therefore, in this paper, we instead start with fine-tuning, one of the most common and easy-to-deploy backdoor defenses, through comprehensive evaluations against diverse attack scenarios. Observations made through initial experiments show that in contrast to the promising defensive results on high poisoning rates, vanilla tuning methods completely fail at low poisoning rate scenarios. Our analysis shows that with the low poisoning rate, the entanglement between backdoor and clean features undermines the effect of tuning-based defenses. Therefore, it is necessary to disentangle the backdoor and clean features in order to improve backdoor purification. To address this, we introduce Feature Shift Tuning (FST), a method for tuning-based backdoor purification. Specifically, FST encourages feature shifts by actively deviating the classifier weights from the originally compromised weights. Extensive experiments demonstrate that our FST provides consistently stable performance under different attack settings. Without complex parameter adjustments, FST also achieves much lower tuning costs, only $10$ epochs. Our codes are available at https://github.com/AISafety-HKUST/stable_backdoor_purification.

Federated learning (FL) is a distributed paradigm that coordinates massive local clients to collaboratively train a global model via stage-wise local training processes on the heterogeneous dataset. Previous works have implicitly studied that FL suffers from the "client-drift" problem, which is caused by the inconsistent optimum across local clients. However, till now it still lacks solid theoretical analysis to explain the impact of this local inconsistency. To alleviate the negative impact of the "client drift" and explore its substance in FL, in this paper, we first design an efficient FL algorithm FedInit, which allows employing the personalized relaxed initialization state at the beginning of each local training stage. Specifically, FedInit initializes the local state by moving away from the current global state towards the reverse direction of the latest local state. This relaxed initialization helps to revise the local divergence and enhance the local consistency level. Moreover, to further understand how inconsistency disrupts performance in FL, we introduce the excess risk analysis and study the divergence term to investigate the test error of the proposed FedInit method. Our studies show that on the non-convex objectives, optimization error is not sensitive to this local inconsistency, while it mainly affects the generalization error bound in FedInit. Extensive experiments are conducted to validate this conclusion. Our proposed FedInit could achieve state-of-the-art (SOTA) results compared to several advanced benchmarks without any additional costs. Meanwhile, stage-wise relaxed initialization could also be incorporated into the current advanced algorithms to achieve higher performance in the FL paradigm.

Data-Free Model Extraction (DFME) aims to clone a black-box model without knowing its original training data distribution, making it much easier for attackers to steal commercial models. Defense against DFME faces several challenges: (i) effectiveness; (ii) efficiency; (iii) no prior on the attacker's query data distribution and strategy. However, existing defense methods: (1) are highly computation and memory inefficient; or (2) need strong assumptions about attack data distribution; or (3) can only delay the attack or prove a model theft after the model stealing has happened. In this work, we propose a Memory and Computation efficient defense approach, named MeCo, to prevent DFME from happening while maintaining the model utility simultaneously by distributionally robust defensive training on the target victim model. Specifically, we randomize the input so that it: (1) causes a mismatch of the knowledge distillation loss for attackers; (2) disturbs the zeroth-order gradient estimation; (3) changes the label prediction for the attack query data. Therefore, the attacker can only extract misleading information from the black-box model. Extensive experiments on defending against both decision-based and score-based DFME demonstrate that MeCo can significantly reduce the effectiveness of existing DFME methods and substantially improve running efficiency.

Dataset condensation (DC) distills a large real-world dataset into a small synthetic dataset, with the goal of training a network from scratch on the latter that performs similarly to the former. State-of-the-art (SOTA) DC methods have achieved satisfactory results through techniques such as accuracy, gradient, training trajectory, or distribution matching. However, these works all perform matching in the high-dimension pixel spaces, ignoring that natural images are usually locally connected and have lower intrinsic dimensions, resulting in low condensation efficiency. In this work, we propose a simple-yet-efficient dataset condensation plugin that matches the raw and synthetic datasets in a low-dimensional manifold. Specifically, our plugin condenses raw images into two low-rank matrices instead of parameterized image matrices. Our plugin can be easily incorporated into existing DC methods, thereby containing richer raw dataset information at limited storage costs to improve the downstream applications' performance. We verify on multiple public datasets that when the proposed plugin is combined with SOTA DC methods, the performance of the network trained on synthetic data is significantly improved compared to traditional DC methods. Moreover, when applying the DC methods as a plugin to continual learning tasks, we observed that our approach effectively mitigates catastrophic forgetting of old tasks under limited memory buffer constraints and avoids the problem of raw data privacy leakage.

The growing size of available data has attracted increasing interest in solving minimax problems in a decentralized manner for various machine learning tasks. Previous theoretical research has primarily focused on the convergence rate and communication complexity of decentralized minimax algorithms, with little attention given to their generalization. In this paper, we investigate the primal-dual generalization bound of the decentralized stochastic gradient descent ascent (D-SGDA) algorithm using the approach of algorithmic stability under both convex-concave and nonconvex-nonconcave settings. Our theory refines the algorithmic stability in a decentralized manner and demonstrates that the decentralized structure does not destroy the stability and generalization of D-SGDA, implying that it can generalize as well as the vanilla SGDA in certain situations. Our results analyze the impact of different topologies on the generalization bound of the D-SGDA algorithm beyond trivial factors such as sample sizes, learning rates, and iterations. We also evaluate the optimization error and balance it with the generalization gap to obtain the optimal population risk of D-SGDA in the convex-concave setting. Additionally, we perform several numerical experiments which validate our theoretical findings.

Existing safe imitation learning (safe IL) methods mainly focus on learning safe policies that are similar to expert ones, but may fail in applications requiring different safety constraints. In this paper, we propose the Lagrangian Generative Adversarial Imitation Learning (LGAIL) algorithm, which can adaptively learn safe policies from a single expert dataset under diverse prescribed safety constraints. To achieve this, we augment GAIL with safety constraints and then relax it as an unconstrained optimization problem by utilizing a Lagrange multiplier. The Lagrange multiplier enables explicit consideration of the safety and is dynamically adjusted to balance the imitation and safety performance during training. Then, we apply a two-stage optimization framework to solve LGAIL: (1) a discriminator is optimized to measure the similarity between the agent-generated data and the expert ones; (2) forward reinforcement learning is employed to improve the similarity while considering safety concerns enabled by a Lagrange multiplier. Furthermore, theoretical analyses on the convergence and safety of LGAIL demonstrate its capability of adaptively learning a safe policy given prescribed safety constraints. At last, extensive experiments in OpenAI Safety Gym conclude the effectiveness of our approach.

We introduce CAMRL, the first curriculum-based asymmetric multi-task learning (AMTL) algorithm for dealing with multiple reinforcement learning (RL) tasks altogether. To mitigate the negative influence of customizing the one-off training order in curriculum-based AMTL, CAMRL switches its training mode between parallel single-task RL and asymmetric multi-task RL (MTRL), according to an indicator regarding the training time, the overall performance, and the performance gap among tasks. To leverage the multi-sourced prior knowledge flexibly and to reduce negative transfer in AMTL, we customize a composite loss with multiple differentiable ranking functions and optimize the loss through alternating optimization and the Frank-Wolfe algorithm. The uncertainty-based automatic adjustment of hyper-parameters is also applied to eliminate the need of laborious hyper-parameter analysis during optimization. By optimizing the composite loss, CAMRL predicts the next training task and continuously revisits the transfer matrix and network weights. We have conducted experiments on a wide range of benchmarks in multi-task RL, covering Gym-minigrid, Meta-world, Atari video games, vision-based PyBullet tasks, and RLBench, to show the improvements of CAMRL over the corresponding single-task RL algorithm and state-of-the-art MTRL algorithms. The code is available at: https://github.com/huanghanchi/CAMRL.

Adaptive optimization has achieved notable success for distributed learning while extending adaptive optimizer to federated Learning (FL) suffers from severe inefficiency, including (i) rugged convergence due to inaccurate gradient estimation in global adaptive optimizer; (ii) client drifts exacerbated by local over-fitting with the local adaptive optimizer. In this work, we propose a novel momentum-based algorithm via utilizing the global gradient descent and locally adaptive amended optimizer to tackle these difficulties. Specifically, we incorporate a locally amended technique to the adaptive optimizer, named Federated Local ADaptive Amended optimizer (FedLADA), which estimates the global average offset in the previous communication round and corrects the local offset through a momentum-like term to further improve the empirical training speed and mitigate the heterogeneous over-fitting. Theoretically, we establish the convergence rate of FedLADA with a linear speedup property on the non-convex case under the partial participation settings. Moreover, we conduct extensive experiments on the real-world dataset to demonstrate the efficacy of our proposed FedLADA, which could greatly reduce the communication rounds and achieves higher accuracy than several baselines.

Continual learning (CL) aims to learn a non-stationary data distribution and not forget previous knowledge. The effectiveness of existing approaches that rely on memory replay can decrease over time as the model tends to overfit the stored examples. As a result, the model's ability to generalize well is significantly constrained. Additionally, these methods often overlook the inherent uncertainty in the memory data distribution, which differs significantly from the distribution of all previous data examples. To overcome these issues, we propose a principled memory evolution framework that dynamically adjusts the memory data distribution. This evolution is achieved by employing distributionally robust optimization (DRO) to make the memory buffer increasingly difficult to memorize. We consider two types of constraints in DRO: f-divergence and Wasserstein ball constraints. For f-divergence constraint, we derive a family of methods to evolve the memory buffer data in the continuous probability measure space with Wasserstein gradient flow (WGF). For Wasserstein ball constraint, we directly solve it in the euclidean space. Extensive experiments on existing benchmarks demonstrate the effectiveness of the proposed methods for alleviating forgetting. As a by-product of the proposed framework, our method is more robust to adversarial examples than compared CL methods.

Few-shot classification is a challenging task of computer vision and is critical to the data-sparse scenario like rare disease diagnosis. Feature augmentation is a straightforward way to alleviate the data-sparse issue in few-shot classification. However, mimicking the original feature distribution from a small amount of data is challenging. Existing augmentation-based methods are task-agnostic: the augmented feature is not with optimal intra-class diversity and inter-class discriminability concerning a certain task. To address this drawback, we propose a novel Task-adaptive Feature Disentanglement and Hallucination framework, dubbed TaFDH. Concretely, we first perceive the task information to disentangle the original feature into two components: class-irrelevant and class-specific features. Then more class-irrelevant features are decoded from a learned variational distribution, fused with the class-specific feature to get the augmented features. Finally, a generalized prior distribution over a quadratic classifier is meta-learned, which can be fast adapted to the class-specific posterior, thus further alleviating the inadequacy and uncertainty of feature hallucination via the nature of Bayesian inference. In this way, we construct a more discriminable embedding space with reasonable intra-class diversity instead of simply restoring the original embedding space, which can lead to a more precise decision boundary. We obtain the augmented features equipped with enhanced inter-class discriminability by highlighting the most discriminable part while boosting the intra-class diversity by fusing with the diverse generated class-irrelevant parts. Experiments on five multi-grained few-shot classification datasets demonstrate the superiority of our method.

Single Image Super-Resolution (SISR) tasks have achieved significant performance with deep neural networks. However, the large number of parameters in CNN-based methods for SISR tasks require heavy computations. Although several efficient SISR models have been recently proposed, most are handcrafted and thus lack flexibility. In this work, we propose a novel differentiable Neural Architecture Search (NAS) approach on both the cell-level and network-level to search for lightweight SISR models. Specifically, the cell-level search space is designed based on an information distillation mechanism, focusing on the combinations of lightweight operations and aiming to build a more lightweight and accurate SR structure. The network-level search space is designed to consider the feature connections among the cells and aims to find which information flow benefits the cell most to boost the performance. Unlike the existing Reinforcement Learning (RL) or Evolutionary Algorithm (EA) based NAS methods for SISR tasks, our search pipeline is fully differentiable, and the lightweight SISR models can be efficiently searched on both the cell-level and network-level jointly on a single GPU. Experiments show that our methods can achieve state-of-the-art performance on the benchmark datasets in terms of PSNR, SSIM, and model complexity with merely 68G Multi-Adds for$\times 2$and 18G Multi-Adds for$\times 4$SR tasks.

Graph classification is a fundamental problem with diverse applications in bioinformatics and chemistry. Due to the intricate procedures of manual annotations in graphical domains, there may be abundant noisy labels of graphs in practice, resulting in poor performance for existing supervised methods. Thus, it is necessary and urgent to study the problem of graph classification with label noise. However, this problem is challenging due to the overfitting of noisy data as well as complicated relational structures of graphs. To handle this problem, we present a simple but effective approach called cOupledMix forGraph Contrast (OMG), which combines coupled Mixup with graph contrastive learning in the feature space. On the one hand, to improve the model generalization, we take convex combination of sample pairs in the feature space for positive pair construction. On the other hand, to accomplish effective optimization, we offer challenging negatives by multiple sample Mixup with different emphasis. To further reduce the impact of noisy data, we develop a neighbour-aware noise removal strategy, which promotes the smoothness in the neighbourhood of samples following the principle of curriculum learning. Extensive experiments on a range of benchmark datasets demonstrate the superiority of our proposed OMG.

Although the vision-and-language pretraining (VLP) equipped cross-modal image-text retrieval (ITR) has achieved remarkable progress in the past two years, it suffers from a major drawback: the ever-increasing size of VLP models restrict its deployment to real-world search scenarios (where the high latency is unacceptable). To alleviate this problem, we present a novel plug-in dynamic contrastive distillation (DCD) framework to compress the large VLP models for the ITR task. Technically, we face the following two challenges: 1) the typical uni-modal metric learning approach is difficult to directly apply to cross-modal task, due to the limited GPU memory to optimize too many negative samples during handling cross-modal fusion features. 2) it is inefficient to static optimize the student network from different hard samples, which have different effects on distillation learning and student network optimization. We try to overcome these challenges from two points. First, to achieve multi-modal contrastive learning, and balance the training costs and effects, we propose to use a teacher network to estimate the difficult samples for students, making the students absorb the powerful knowledge from pre-trained teachers, and master the knowledge from hard samples. Second, to dynamic learn from hard sample pairs, we propose dynamic distillation to dynamically learn samples of different difficulties, from the perspective of better balancing the difficulty of knowledge and students' self-learning ability. We successfully apply our proposed DCD strategy on two state-of-the-art vision-language pretrained models, i.e. ViLT and METER. Extensive experiments on MS-COCO and Flickr 30 K benchmarks show the effectiveness and efficiency of our DCD framework. Encouragingly, we can speed up the inference at least 129 × compared to the existing ITR models. We further provide in-depth analyses and discussions that explain where the performance improvement comes from. We hope our work can shed light on other tasks that require distillation and contrastive learning.

Pre-Training (PT) of text representations has been successfully applied to low-resource Neural Machine Translation (NMT). However, it usually fails to achieve notable gains (some- times, even worse) on resource-rich NMT on par with its Random-Initialization (RI) counterpart. We take the first step to investigate the complementarity between PT and RI in resource-rich scenarios via two probing analyses, and find that: 1) PT improves NOT the accuracy, but the generalization by achieving flatter loss landscapes than that of RI; 2) PT improves NOT the confidence of lexical choice, but the negative diversity by assigning smoother lexical probability distributions than that of RI. Based on these insights, we propose to combine their complementarities with a model fusion algorithm that utilizes optimal transport to align neurons between PT and RI. Experiments on two resource-rich translation benchmarks, WMT’17 English-Chinese (20M) and WMT’19 English-German (36M), show that PT and RI could be nicely complementary to each other, achieving substantial improvements considering both translation accuracy, generalization, and negative diversity. Probing tools and code are released at: https://github.com/zanchangtong/PTvsRI.

Catastrophic forgetting (CF) frequently occurs when learning with non-stationary data distribution. The CF issue remains nearly unexplored and is more challenging when meta-learning on a sequence of domains (datasets), called sequential domain meta-learning (SDML). In this work, we propose a simple yet effective learning to learn approach, i.e., meta optimizer, to mitigate the CF problem in SDML. We first apply the proposed meta optimizer to the simplified setting of SDML, domain-aware meta-learning, where the domain labels and boundaries are known during the learning process. We propose dynamically freezing the network and incorporating it with the proposed meta optimizer by considering the domain nature during meta training. In addition, we extend the meta optimizer to the more general setting of SDML, domain-agnostic meta-learning, where domain labels and boundaries are unknown during the learning process. We propose a domain shift detection technique to capture latent domain change and equip the meta optimizer with it to work in this setting. The proposed meta optimizer is versatile and can be easily integrated with several existing meta-learning algorithms. Finally, we construct a challenging and large-scale benchmark consisting of 10 heterogeneous domains with a super long task sequence consisting of 100K tasks. We perform extensive experiments on the proposed benchmark for both settings and demonstrate the effectiveness of our proposed method, outperforming current strong baselines by a large margin.

Federated Learning (FL) is an emerging distributed learning paradigm under privacy constraint. Data heterogeneity is one of the main challenges in FL, which results in slow convergence and degraded performance. Most existing approaches only tackle the heterogeneity challenge by restricting the local model update in client, ignoring the performance drop caused by direct global model aggregation. Instead, we propose a data-free knowledge distillation method to fine-tune the global model in the server (FedFTG), which relieves the issue of direct model aggregation. Concretely, FedFTG explores the input space of local models through a generator, and uses it to transfer the knowledge from local models to the global model. Besides, we propose a hard sample mining scheme to achieve effective knowledge distillation throughout the training. In addition, we develop customized label sampling and class-level ensemble to derive maximum utilization of knowledge, which implicitly mitigates the distribution discrepancy across clients. Extensive experiments show that our FedFTG significantly outperforms the state-of-the-art (SOTA) FL algorithms and can serve as a strong plugin for enhancing FedAvg, FedProx, FedDyn, and SCAFFOLD.

Personalized federated learning is proposed to handle the data heterogeneity problem amongst clients by learning dedicated tailored local models for each user. However, existing works are often built in a centralized way, leading to high communication pressure and high vulnerability when a failure or an attack on the central server occurs. In this work, we propose a novel personalized federated learning framework in a decentralized (peer-to-peer) communication protocol named DisPFL, which employs personalized sparse masks to customize sparse local models on the edge. To further save the communication and computation cost, we propose a decentralized sparse training technique, which means that each local model in DisPFL only maintains a fixed number of active parameters throughout the whole local training and peer-to-peer communication process. Comprehensive experiments demonstrate that DisPFL significantly saves the communication bottleneck for the busiest node among all clients and, at the same time, achieves higher model accuracy with less computation cost and communication rounds. Furthermore, we demonstrate that our method can easily adapt to heterogeneous local clients with varying computation complexities and achieves better personalized performances.

Model fusion without accessing training data in machine learning has attracted increasing interest due to the practical resource-saving and data privacy issues. During the training process, the neural weights of each model can be randomly permuted, and we have to align the channels of each layer before fusing them. Regrading the channels as nodes and weights as edges, aligning the channels to maximize weight similarity is a challenging NP-hard assignment problem. Due to its quadratic assignment nature, we formulate the model fusion problem as a graph matching task, considering the second-order similarity of model weights instead of previous work merely formulating model fusion as a linear assignment problem. For the rising problem scale and multi-model consistency issues, we propose an efficient graduated assignment-based model fusion method, dubbed GAMF, which iteratively updates the matchings in a consistency-maintaining manner. We apply GAMF to tackle the compact model ensemble task and federated learning task on MNIST, CIFAR-10, CIFAR-100, and Tiny-Imagenet. The performance shows the efficacy of our GAMF compared to state-of-the-art baselines.

Task-free continual learning (CL) aims to learn a non-stationary data stream without explicit task definitions and not forget previous knowledge. The widely adopted memory replay approach could gradually become less effective for long data streams, as the model may memorize the stored examples and overfit the memory buffer. Second, existing methods overlook the high uncertainty in the memory data distribution since there is a big gap between the memory data distribution and the distribution of all the previous data examples. To address these problems, for the first time, we propose a principled memory evolution framework to dynamically evolve the memory data distribution by making the memory buffer gradually harder to be memorized with distributionally robust optimization (DRO). We then derive a family of methods to evolve the memory buffer data in the continuous probability measure space with Wasserstein gradient flow (WGF). The proposed DRO is w.r.t the worst-case evolved memory data distribution, thus guarantees the model performance and learns significantly more robust features than existing memory-replay-based methods. Extensive experiments on existing benchmarks demonstrate the effectiveness of the proposed methods for alleviating forgetting. As a by-product of the proposed framework, our method is more robust to adversarial examples than existing task-free CL methods.

Robust overfitting widely exists in adversarial training of deep networks. The exact underlying reasons for this are still not completely understood. Here, we explore the causes of robust overfitting by comparing the data distribution of non-overfit (weak adversary) and overfitted (strong adversary) adversarial training, and observe that the distribution of the adversarial data generated by weak adversary mainly contain small-loss data. However, the adversarial data generated by strong adversary is more diversely distributed on the large-loss data and the small-loss data. Given these observations, we further designed data ablation adversarial training and identify that some small-loss data which are not worthy of the adversary strength cause robust overfitting in the strong adversary mode. To relieve this issue, we propose minimum loss constrained adversarial training (MLCAT): in a minibatch, we learn large-loss data as usual, and adopt additional measures to increase the loss of the small-loss data. Technically, MLCAT hinders data fitting when they become easy to learn to prevent robust overfitting; philosophically, MLCAT reflects the spirit of turning waste into treasure and making the best use of each adversarial data; algorithmically, we designed two realizations of MLCAT, and extensive experiments demonstrate that MLCAT can eliminate robust overfitting and further boost adversarial robustness.

Overfitting widely exists in adversarial robust training of deep networks. An effective remedy is adversarial weight perturbation, which injects the worst-case weight perturbation during network training by maximizing the classification loss on adversarial examples. Adversarial weight perturbation helps reduce the robust generalization gap; however, it also undermines the robustness improvement. A criterion that regulates the weight perturbation is therefore crucial for adversarial training. In this paper, we propose such a criterion, namely Loss Stationary Condition (LSC) for constrained perturbation. With LSC, we find that it is essential to conduct weight perturbation on adversarial data with small classification loss to eliminate robust overfitting. Weight perturbation on adversarial data with large classification loss is not necessary and may even lead to poor robustness. Based on these observations, we propose a robust perturbation strategy to constrain the extent of weight perturbation. The perturbation strategy prevents deep networks from overfitting while avoiding the side effect of excessive weight perturbation, significantly improving the robustness of adversarial training. Extensive experiments demonstrate the superiority of the proposed method over the state-of-the-art adversarial training methods.

Safe reinforcement learning aims to learn the optimal policy while satisfying safety constraints, which is essential in real-world applications. However, current algorithms still struggle for efficient policy updates with hard constraint satisfaction. In this paper, we propose Penalized Proximal Policy Optimization (P3O), which solves the cumbersome constrained policy iteration via a single minimization of an equivalent unconstrained problem. Specifically, P3O utilizes a simple yet effective penalty approach to eliminate cost constraints and removes the trust-region constraint by the clipped surrogate objective. We theoretically prove the exactness of the penalized method with a finite penalty factor and provide a worst-case analysis for approximate error when evaluated on sample trajectories. Moreover, we extend P3O to more challenging multi-constraint and multi-agent scenarios which are less studied in previous work. Extensive experiments show that P3O outperforms state-of-the-art algorithms with respect to both reward improvement and constraint satisfaction on a set of constrained locomotive tasks.

Learning a risk-aware policy is essential but rather challenging in unstructured robotic tasks. Safe reinforcement learning methods open up new possibilities to tackle this problem. However, the conservative policy updates make it intractable to achieve sufficient exploration and desirable performance in complex, sample-expensive environments. In this paper, we propose a dual-agent safe reinforcement learning strategy consisting of a baseline and a safe agent. Such a decoupled framework enables high flexibility, data efficiency and risk-awareness for RL-based control. Concretely, the baseline agent is responsible for maximizing rewards under standard RL settings. Thus, it is compatible with off-the-shelf training techniques of unconstrained optimization, exploration and exploitation. On the other hand, the safe agent mimics the baseline agent for policy improvement and learns to fulfill safety constraints via off-policy RL tuning. In contrast to training from scratch, safe policy correction requires significantly fewer interactions to obtain a near-optimal policy. The dual policies can be optimized synchronously via a shared replay buffer, or leveraging the pre-trained model or the non-learning-based controller as a fixed baseline agent. Experimental results show that our approach can learn feasible skills without prior knowledge as well as deriving risk-averse counterparts from pre-trained unsafe policies. The proposed method outperforms the state-of-the-art safe RL algorithms on difficult robot locomotion and manipulation tasks with respect to both safety constraint satisfaction and sample efficiency.

Graph neural networks (GNNs) have achieved state-of-the-art results on graph classification tasks. They have been primarily studied in cases of supervised end-to-end training, which requires abundant task-specific labels. Unfortunately, annotating labels of graph data could be prohibitively expensive or even impossible in many applications. An effective solution is to incorporate labeled graphs from a different, but related source domain, to develop a graph classification model for the target domain. However, the problem of unsupervised domain adaptation for graph classification is challenging due to potential domain discrepancy in graph space as well as the label scarcity in the target domain. In this paper, we present a novel GNN framework named DEAL by incorporating both source graphs and target graphs, which is featured by two modules, i.e., adversarial perturbation and pseudo-label distilling. Specifically, to overcome domain discrepancy, we equip source graphs with target semantics by applying to them adaptive perturbations which are adversarially trained against a domain discriminator. Additionally, DEAL explores distinct feature spaces at different layers of the GNN encoder, which emphasize global and local semantics respectively. Then, we distill the consistent predictions from two spaces to generate reliable pseudo-labels for sufficiently utilizing unlabeled data, which further improves the performance of graph classification. Extensive experiments on a wide range of graph classification datasets reveal the effectiveness of our proposed DEAL.

State-of-the-art causal discovery methods usually assume that the observational data is complete. However, the missing data problem is pervasive in many practical scenarios such as clinical trials, economics, and biology. One straightforward way to address the missing data problem is first to impute the data using off-the-shelf imputation methods and then apply existing causal discovery methods. However, such a two-step method may suffer from suboptimality, as the imputation algorithm may introduce bias for modeling the underlying data distribution. In this paper, we develop a general method, which we call MissDAG, to perform causal discovery from data with incomplete observations. Focusing mainly on the assumptions of ignorable missingness and the identifiable additive noise models (ANMs), MissDAG maximizes the expected likelihood of the visible part of observations under the expectation-maximization (EM) framework. In the E-step, in cases where computing the posterior distributions of parameters in closed-form is not feasible, Monte Carlo EM is leveraged to approximate the likelihood. In the M-step, MissDAG leverages the density transformation to model the noise distributions with simpler and specific formulations by virtue of the ANMs and uses a likelihood-based causal discovery algorithm with directed acyclic graph constraint. We demonstrate the flexibility of MissDAG for incorporating various causal discovery algorithms and its efficacy through extensive simulations and real data experiments.

Deep neural networks often suffer from poor generalization caused by complex and non-convex loss landscapes. One of the popular solutions is Sharpness-Aware Minimization (SAM), which smooths the loss landscape via minimizing the maximized change of training loss when adding a perturbation to the weight. However, we find the indiscriminate perturbation of SAM on all parameters is suboptimal, which also results in excessive computation,~\emph{i.e.}, double the overhead of common optimizers like Stochastic Gradient Descent~(SGD). In this paper, we propose an efficient and effective training scheme coined as Sparse SAM (SSAM), which achieves sparse perturbation by a binary mask. To obtain the sparse mask, we provide two solutions which are based on Fisher information and dynamic sparse training, respectively. In addition, we theoretically prove that SSAM can converge at the same rate as SAM,~\emph{i.e.}, $O(\log T/\sqrt{T})$. Sparse SAM not only has the potential for training acceleration but also smooths the loss landscape effectively. Extensive experimental results on CIFAR10, CIFAR100, and ImageNet-1K confirm the superior efficiency of our method to SAM, and the performance is preserved or even better with a perturbation of merely 50\% sparsity. Code is available at \url{https://github.com/Mi-Peng/Sparse-Sharpness-Aware-Minimization}.

Deep neural networks (DNNs) have been shown to be vulnerable to adversarial examples, which can produce erroneous predictions by injecting imperceptible perturbations. In this work, we study the transferability of adversarial examples, which is significant due to its threat to real-world applications where model architecture or parameters are usually unknown. Many existing works reveal that the adversarial examples are likely to overfit the surrogate model that they are generated from, limiting its transfer attack performance against different target models. To mitigate the overfitting of the surrogate model, we propose a novel attack method, dubbed reverse adversarial perturbation (RAP). Specifically, instead of minimizing the loss of a single adversarial point, we advocate seeking adversarial example located at a region with unified low loss value, by injecting the worst-case perturbation (the reverse adversarial perturbation) for each step of the optimization procedure. The adversarial attack with RAP is formulated as a min-max bi-level optimization problem. By integrating RAP into the iterative process for attacks, our method can find more stable adversarial examples which are less sensitive to the changes of decision boundary, mitigating the overfitting of the surrogate model. Comprehensive experimental comparisons demonstrate that RAP can significantly boost adversarial transferability. Furthermore, RAP can be naturally combined with many existing black-box attack techniques, to further boost the transferability. When attacking a real-world image recognition system, Google Cloud Vision API, we obtain 22% performance improvement of targeted attacks over the compared method. Our codes are available at https://github.com/SCLBD/TransferattackRAP.

Deep neural networks (DNN) based recommender models often require numerous parameters to achieve remarkable performance. However, this inevitably brings redundant neurons, a phenomenon referred to as over-parameterization. In this paper, we plan to exploit such redundancy phenomena for recommender systems (RS), and propose a top-N item recommendation framework called PCRec that leverages collaborative training of two recommender models of the same network structure, termed peer collaboration. We first introduce two criteria to identify the importance of parameters of a given recommender model. Then, we rejuvenate the unimportant parameters by copying parameters from its peer network. After such an operation and retraining, the original recommender model is endowed with more representation capacity by possessing more functional model parameters. To show its generality, we instantiate PCRec by using three well-known recommender models. We conduct extensive experiments on two real-world datasets, and show that PCRec yields significantly better performance than its counterpart with the same model (parameter) size.

Existing meta-learning works assume that each task has available training and testing data. However, there are many available pre-trained models without accessing their training data in practice. We often need a single model to solve different tasks simultaneously as this is much more convenient to deploy the models. Our work aims to meta-learn a model initialization from these pre-trained models without using corresponding training data. We name this challenging problem setting as Data-Free Learning To Learn (DFL2L). We propose a distributionally robust optimization (DRO) framework to learn a black-box model to fuse and compress all the pre-trained models into a single network to address this problem. To encourage good generalization to the unseen new tasks, the proposed DRO framework diversifies the learned task embedding associated with each pre-trained model to cover the diversity in the underlying training task distributions. A model initialization is sampled from the black-box network during meta-testing as the meta learned initialization. Extensive experiments on offline and online DFL2L settings and several real image datasets demonstrate the effectiveness of the proposed methods.

Federated learning (FL), arising as a privacy-preserving machine learning paradigm, has received notable attention from the public. In each round of synchronous FL training, only a fraction of available clients are chosen to participate, and the selection decision might have a significant effect on the training efficiency, as well as the final model performance. In this article, we investigate the client selection problem under a volatile context, in which the local training of heterogeneous clients is likely to fail due to various kinds of reasons and in different levels of frequency. Intuitively, too much training failure might potentially reduce the training efficiency, while too much selection on clients with greater stability might introduce bias, thereby resulting in degradation of the training effectiveness. To tackle this tradeoff, we, in this article, formulate the client selection problem under joint consideration of effective participation and fairness. Furthermore, we propose E3CS, a stochastic client selection scheme as a solution. According to our experimental results over a public data set, the proposed selection scheme is able to achieve up to$2\times $faster convergence to a fixed model accuracy while maintaining the same level of final model accuracy, compared with the state-of-the-art selection schemes.

Generative Adversarial Networks (GANs) are formulated as minimax game problems that generative networks attempt to approach real data distributions by adversarial learning against discriminators which learn to distinguish generated samples from real ones, of which the intrinsic problem complexity poses challenges to performance and robustness. In this work, we aim to boost model learning from the perspective of network architectures, by incorporating recent progress on automated architecture search into GANs. Specially we propose a fully differentiable search framework, dubbedalphaGAN, where the searching process is formalized as solving a bi-level minimax optimization problem. The outer-level objective aims for seeking an optimal network architecture towards pure Nash Equilibrium conditioned on the network parameters of generators and discriminators optimized with a traditional adversarial loss within inner level. The entire optimization performs a first-order approach by alternately minimizing the two-level objective in a fully differentiable manner that enables obtaining a suitable architecture efficiently from an enormous search space. Extensive experiments on CIFAR-10 and STL-10 datasets show that our algorithm can obtain high-performing architectures only with 3-GPU hours on a single GPU in the search space comprised of approximate$2\times 10^{11}$possible configurations. We further validate the method on the state-of-the-art network StyleGAN2, and push the score of Fréchet Inception Distance (FID) further, i.e., achieving 1.94 on CelebA, 2.86 on LSUN-church and 2.75 on FFHQ, with relative improvements$3\%{\sim} 26\%$over the baseline architecture. We also provide a comprehensive analysis of the behavior of the searching process and the properties of searched architectures, which would benefit further research on architectures for generative models. Codes and models are available athttps://github.com/yuesongtian/AlphaGAN.

Learning Directed Acyclic Graph (DAG) from purely observational data is a critical problem for causal inference. Most existing works tackle this problem by exploring gradient-based learning methods with a smooth characterization of acyclicity. A major shortcoming of current gradient based works is that they independently optimize SEMs with a single sample and neglect the interactions between different samples. In this paper, we consider DAG structure learning from the perspective of distributional optimization and design an adversarial framework named DAG-GAN to detect the DAG structure from data. We theoretically analyze the Nash equilibrium property of DAG-GAN and propose a novel score function to exploit the interactions between different samples. In addition, extensive experiments are conducted to validate the efficiency of DAG-GAN against several state-of-the-art DAG learning methods.

Works on lottery ticket hypothesis (LTH) and single-shot network pruning (SNIP) have raised a lot of attention currently on post-training pruning (iterative magnitude pruning), and before-training pruning (pruning at initialization). The former method suffers from an extremely large computation cost and the latter usually struggles with insufficient performance. In comparison, during-training pruning, a class of pruning methods that simultaneously enjoys the training/inference efficiency and the comparable performance, temporarily, has been less explored. To better understand during-training pruning, we quantitatively study the effect of pruning throughout training from the perspective of pruning plasticity (the ability of the pruned networks to recover the original performance). Pruning plasticity can help explain several other empirical observations about neural network pruning in literature. We further find that pruning plasticity can be substantially improved by injecting a brain-inspired mechanism called neuroregeneration, i.e., to regenerate the same number of connections as pruned. We design a novel gradual magnitude pruning (GMP) method, named gradual pruning with zero-cost neuroregeneration (GraNet), that advances state of the art. Perhaps most impressively, its sparse-to-sparse version for the first time boosts the sparse-to-sparse training performance over various dense-to-sparse methods with ResNet-50 on ImageNet without extending the training time. We release all codes in https://github.com/Shiweiliuiiiiiii/GraNet.

Knowledge distillation has proven to be an effective model compression method that exploits the knowledge from a teacher model for supervising a student model by minimizing the distribution difference between the knowledge and the prediction produced by the student model. In this work, we focus on improving its performance from the perspective of divergence measures. A general form representing a family of divergence measures is introduced and a novel learning paradigm that jointly optimizes multiple measures is proposed by formalizing it as a multi-objective learning problem. Conditioned on Pareto optimality, the weights of different divergences are tuned in an automated way during training. Extensive experiments on multiple datasets show the proposed method can significantly improve the performance of student networks compared to the state-of-the-art methods for knowledge distillation. Codes are available at https://github.com/CML00/MoDiv.

Recent research has witnessed advances in facial image editing tasks including face swapping and face reenactment. However, these methods are confined to dealing with one specific task at a time. In addition, for video facial editing, previous methods either simply apply transformations frame by frame or utilize multiple frames in a concatenated or iterative fashion, which leads to noticeable visual flickers. In this paper, we propose a unified temporally consistent facial video editing framework termed UniFaceGAN. Based on a 3D reconstruction model and a simple yet efficient dynamic training sample selection mechanism, our framework is designed to handle face swapping and face reenactment simultaneously. To enforce the temporal consistency, a novel 3D temporal loss constraint is introduced based on the barycentric coordinate interpolation. Besides, we propose a region-aware conditional normalization layer to replace the traditional AdaIN or SPADE to synthesize more context-harmonious results. Compared with the state-of-the-art facial image editing methods, our framework generates video portraits that are more photo-realistic and temporally smooth.

In this article, we present a distributed variant of an adaptive stochastic gradient method for training deep neural networks in the parameter-server model. To reduce the communication cost among the workers and server, we incorporate two types of quantization schemes, i.e., gradient quantization and weight quantization, into the proposed distributed Adam. In addition, to reduce the bias introduced by quantization operations, we propose an error-feedback technique to compensate for the quantized gradient. Theoretically, in the stochastic nonconvex setting, we show that the distributed adaptive gradient method with gradient quantization and error feedback converges to the first-order stationary point, and that the distributed adaptive gradient method with weight quantization and error feedback converges to the point related to the quantized level under both the single-worker and multi-worker modes. Last, we apply the proposed distributed adaptive gradient methods to train deep neural networks. Experimental results demonstrate the efficacy of our methods.

Multi-task learning (MTL) is a common paradigm that seeks to improve the generalization performance of task learning by training related tasks simultaneously. However, it is still a challenging problem to search the flexible and accurate architecture that can be shared among multiple tasks. In this paper, we propose a novel deep learning model called Task Adaptive Activation Network (TAAN) that can automatically learn the optimal network architecture for MTL. The main principle of TAAN is to derive flexible activation functions for different tasks from the data with other parameters of the network fully shared. We further propose two functional regularization methods that improve the MTL performance of TAAN. The improved performance of both TAAN and the regularization methods is demonstrated by comprehensive experiments.

Sparse optimization is a central problem in machine learning and computer vision. However, this problem is inherently NP-hard and thus difficult to solve in general. Combinatorial search methods find the global optimal solution but are confined to small-sized problems, while coordinate descent methods are efficient but often suffer from poor local minima. This paper considers a new block decomposition algorithm that combines the effectiveness of combinatorial search methods and the efficiency of coordinate descent methods. Specifically, we consider a random strategy or/and a greedy strategy to select a subset of coordinates as the working set, and then perform a global combinatorial search over the working set based on the original objective function. We show that our method finds stronger stationary points than Amir Beck et al.'s coordinate-wise optimization method. In addition, we establish the convergence rate of our algorithm. Our experiments on solving sparse regularized and sparsity constrained least squares optimization problems demonstrate that our method achieves state-of-the-art performance in terms of accuracy. For example, our method generally outperforms the well-known greedy pursuit method.

The sparse generalized eigenvalue problem arises in a number of standard and modern statistical learning models, including sparse principal component analysis, sparse Fisher discriminant analysis, and sparse canonical correlation analysis. However, this problem is difficult to solve since it is NP-hard. In this paper, we consider a new effective decomposition method to tackle this problem. Specifically, we use random or/and swapping strategies to find a working set and perform global combinatorial search over the small subset of variables. We consider a bisection search method and a coordinate descent method for solving the quadratic fractional programming subproblem. In addition, we provide some theoretical analysis for the proposed method. Our experiments on synthetic data and real-world data have shown that our method significantly and consistently outperforms existing solutions in term of accuracy.

Trust-aware recommender systems have received much attention recently for their abilities to capture the influence among connected users. However, they suffer from the efficiency issue due to large amount of data and time-consuming real-valued operations. Although existing discrete collaborative filtering may alleviate this issue to some extent, it is unable to accommodate social influence. In this paper we propose a discrete trust-aware matrix factorization (DTMF) model to take dual advantages of both social relations and discrete technique for fast recommendation. Specifically, we map the latent representation of users and items into a joint hamming space by recovering the rating and trust interactions between users and items. We adopt a sophisticated discrete coordinate descent (DCD) approach to optimize our proposed model. In addition, experiments on two real-world datasets demonstrate the superiority of our approach against other state-of-the-art approaches in terms of ranking accuracy and efficiency.

We propose a fast first-order method to solve multi-term nonsmooth composite convex minimization problems by employing a recent proximal average approximation technique and a novel adaptive parameter tuning technique. Thanks to this powerful parameter tuning technique, the proximal gradient step can be performed with a much larger stepsize in the algorithm implementation compared with the prior PA-APG method, which is the core to enable significant improvements in practical performance. Moreover, by choosing the approximation parameter adaptively, the proposed method is shown to enjoy the O(1/k) iteration complexity theoretically without needing any extra computational cost, while the PA-APG method incurs much more iterations for convergence. The preliminary experimental results on overlapping group Lasso and graph-guided fused Lasso problems confirm our theoretic claim well, and indicate that the proposed method is almost five times faster than the state-of-the-art PA-APG method and therefore suitable for higher-precision required optimization.