Knowledge Distillation (KD) serves as an effective approach to addressing heterogeneity issues in Federated Learning (FL), leveraging additional datasets to align local and global models better. There are two primary distillation paradigms: feature-based distillation, which utilizes intermediate-layer features of the network, and logit-based distillation, which employs the final layer's logit outputs. However, existing studies often select distillation methods based on intuitive and empirical evidence when facing different heterogeneous settings, neglecting the intrinsic relationship between distillation paradigms and heterogeneity. This oversight may result in suboptimal federated knowledge distillation performance under heterogeneous conditions. In this paper, we propose the Consolidated Distillation for Heterogeneous Federated Learning - FedCD that balances knowledge representations from both feature-based and logit-based distillation to enhance performance. Specifically, to address the misalignment between knowledge conveyed by features and logits, we aggregate features from different layers via cross-layer attention to preserve semantic knowledge, followed by distribution modeling using Gaussian Mixture Models. This process strengthens knowledge distillation by constraining the transformation of different network layers' features under a consolidated distribution, thereby mitigating impacts from both data and model heterogeneity. Extensive experiments demonstrate that FedCD outperforms state-of-the-art methods by over 10.72% and validate the effectiveness of our approach.

Data-Centric Sequential Recommendation (DaCSR) has emerged as a promising technique that enhances dataset quality to better capture user preferences without increasing training complexity. However, mining item relations to improve data quality remains challenging due to the intricate nature of interaction sequences. Existing methods predominantly either: 1) optimize models to learn such item relations from fixed datasets at significant training cost, or 2) employ generative models to adaptively learn only interaction patterns, which lack interpretability and cannot guarantee effective data quality enhancement. In this paper, we pioneer a relation-guided dataset augmentation and regeneration framework for sequential recommendation called \textbf{RaSR}. This framework can significantly improve model performance on original datasets while maintaining training efficiency without modifying the model architecture. Specifically, we first preprocess user interactions to construct standardized sequential data and extract semantic representations via a Large Language Model (LLM). We then build a multi-relation graph with manually predefined metrics and semantic representations to generate augmented datasets. Finally, a relation-aware generator can produce regenerated datasets with both the multi-relation graph and the augmented dataset. To verify the effectiveness of RaSR, we conduct experiments on various backbone models and datasets, and achieve significant performance improvement compared to training the model only on the original dataset.

Federated Graph Learning (FGL) has emerged as a powerful paradigm for decentralized training of graph neural networks while preserving data privacy. However, existing FGL methods are predominantly designed for static graphs and rely on parameter averaging or distribution alignment, which implicitly assume that all features are equally transferable across clients, overlooking both the spatial and temporal heterogeneity and the presence of client-specific knowledge in real-world graphs. In this work, we identify that such assumptions create a vicious cycle of spurious representation entanglement, client-specific interference, and negative transfer, degrading generalization performance in Federated Learning over Dynamic Spatio-Temporal Graphs (FSTG). To address this issue, we propose a novel causality-inspired framework named SC-FSGL, which explicitly decouples transferable causal knowledge from client-specific noise through representation-level interventions. Specifically, we introduce a Conditional Separation Module that simulates soft interventions through client conditioned masks, enabling the disentanglement of invariant spatio-temporal causal factors from spurious signals and mitigating representation entanglement caused by client heterogeneity. In addition, we propose a Causal Codebook that clusters causal prototypes and aligns local representations via contrastive learning, promoting cross-client consistency and facilitating knowledge sharing across diverse spatio-temporal patterns. Experiments on five diverse heterogeneity Spatio-Temporal Graph (STG) datasets show that SC-FSGL outperforms state-of-the-art methods.

Fake orders pose increasing threats to sequential recommender systems by misleading recommendation results through artificially manipulated interactions, including click farming, context-irrelevant substitutions, and sequential perturbations. Unlike injecting carefully designed fake users to influence recommendation performance, fake orders embedded within genuine user sequences aim to disrupt user preferences and mislead recommendation results, thereby manipulating exposure rates of specific items to gain competitive advantages. To protect users' authentic interest preferences and eliminate misleading information, this paper aims to perform precise and efficient rectification on compromised sequential recommender systems while avoiding the enormous computational and time costs of retraining existing models. Specifically, we identify that fake orders are not absolutely harmful—in certain cases, partial fake orders can even have a data augmentation effect. Based on this insight, we propose Dual-view Identification and Targeted Rectification (DITaR), which primarily identifies harmful samples to achieve unbiased rectification of the system. The core idea of this method is to obtain differentiated representations from collaborative and semantic views for precise detection, and then filters detected suspicious fake orders to select truly harmful ones for targeted rectification with gradient ascent. This ensures that useful information in fake orders is not removed while preventing bias residue. Moreover, it maintains the original data volume and sequence structure, thus protecting system performance and trustworthiness to achieve optimal unbiased rectification. Extensive experiments on three datasets demonstrate that DITaR achieves superior performance compared to state-of-the-art methods in terms of recommendation quality, computational efficiency, and system robustness.

Contribution evaluation is essential for incentivizing high-quality data sharing in federated learning (FL), yet existing Shapley-value-based methods are prohibitively expensive and overlook temporal influence propagation. In this paper, we propose Ripple Shapley, a novel attribution framework that enables accurate, real-time data valuation within a single federated training run. Our method decomposes each sample’s impact into an instantaneous drop term and a recursive ripple term, the latter capturing downstream influence via a Jacobian chain over global updates. To scale computation, we introduce a low-rank approximation of the Jacobian product and construct a shared subspace for efficient ripple accumulation. Extensive experiments on CIFAR-10 and MNIST show that Ripple Shapley achieves up to 62× speedup over existing Shapley-based FL methods while maintaining high attribution fidelity, significantly improving efficiency, robustness, and fairness in federated environments. We further demonstrate its effectiveness in dynamic federated learning scenarios and its potential for real-time data pricing.

Xiaoquan Yi, Haixing Wu, Haozhao Wang, Yichen Li, Yuhua Li, Rui Zhang, Ruixuan Li. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

The deployment of large language models' (LLMs) inference at the edge can facilitate prompt service responsiveness while protecting user privacy. However, it is critically challenged by the resource constraints of a single edge node. Distributed inference has emerged to aggregate and leverage computational resources across multiple devices. Yet, existing methods typically require strict synchronization, which is often infeasible due to the unreliable network conditions. In this paper, we propose HALO , a novel framework that can boost the distributed LLM inference in lossy edge network. The core idea is to enable a relaxed yet effective synchronization by strategically allocating less critical neuron groups to unstable devices, thus avoiding the excessive waiting time incurred by delayed packets. HALO introduces three key mechanisms: (1) a semantic-aware predictor to assess the significance of neuron groups prior to activation. (2) a parallel execution scheme of neuron group loading during the model inference. (3) a load-balancing scheduler that efficiently orchestrates multiple devices with heterogeneous resources. Experimental results from a Raspberry Pi cluster demonstrate that HALO achieves a 3.41x end-to-end speedup for LLaMA-series LLMs under unreliable network conditions. It maintains performance comparable to optimal conditions and significantly outperforms the state-of-the-art in various scenarios.

Federated recommendation systems (FRSs) have recently gained widespread attention due to their ability to train collaborative recommendation models without exchanging raw user data. However, existing FRSs face a severe challenge of data sparsity, which manifests at both the user and item levels. First, user data sparsity: some users may only have a small number of interactions with items, struggling to adequately train the personalized user embedding locally. Second, item data sparsity: some items may only receive a small number of user ratings, causing the global model to lack knowledge about them. Considering these, we propose the Knowledge Enhanced Federated Recommendation System named as KE-FedRS, of which the core idea is to enhance the knowledge of users with few interactions and items with few ratings at both the local and global levels. Specifically, at the local level, we introduce an auxiliary user embedding and average and aggregate this auxiliary embedding across similar users, thereby enriching the knowledge of the local user embedding. At the global level, we propose a hybrid client selection strategy based on item embedding discrepancies, prioritizing clients that exhibit greater divergence in item embeddings from others, thus enhancing the knowledge of items with fewer interactions in the global model. We conduct comprehensive experiments on four real-world datasets, and the results show that the proposed method consistently outperforms baseline approaches in terms of HR@10 and NDCG@10.

By allowing each client to refer to the knowledge from other clients while retaining their specific characteristics, partial knowledge sharing has become one of the main approaches to realizing personalized federated learning (pFL). Representative techniques of partial knowledge sharing propose sharing the feature extractor while customizing the classifier head of the neural network. Although such methods achieve great success, the underlying principle behind them remains yet to be comprehensively understood. A fundamental problem is whether it is really appropriate to fully share the feature extractor. Based on the theory of neural collapse, in this paper, we demonstrate both theoretically and empirically that the feature extractor should be partially shared rather than fully shared. More specifically, we identify a substantial inconsistency between the fused global feature representations and expected local feature representations, and thus it is necessary to preserve partially customized layers of the feature extractor for enhancing personalized representations. Based on this discovery, we further propose a novel method called pFedDKS which detaches the shared global knowledge and customized local knowledge by providing detached feature prototypes. Extensive experiments on various datasets and models show that pFedDKS outperforms state-of-the-arts.

The sparse mixture-of-experts (MoE) architecture is a promising backbone of foundation models for a wide range of applications in edge. However, deploying them locally presents a significant challenge to memory-constrained GPUs. Previous techniques utilize CPUs for expert offloading, which suffer from inaccurate expert prefetching and on-demand loading latency. To address these challenges, we propose self-speculative MoE (SS-MoE), an algorithm-system co-design framework that facilitates inference under limited GPU memory. Our insight is that only a subset of routed experts, i.e., draft model, can still tackle easy tasks and generate draft tokens. Second, we deem GPU memory as the experts cache, and on-demand update it to mitigate IO overhead. Draft tokens from fewer routed experts are generated quickly, and these experts are then routed for verification. Additionally, we design a confidence-based policy to adaptively accept or verify draft tokens, which selectively decreases or increases the number of verification tokens of speculative decoding and achieves acceleration. Notably, under conservative verification, our approach preserves model accuracy and surpasses the decoding speed of the 4-bit quantized counterpart model. Under adaptive verification, our method significantly enhances decoding speed by 3.72x over state-of-the-art methods while maintaining nearly lossless accuracy.

Accurate dynamic modeling of task correlations is crucial for enhancing collaborative efficiency and personalized performance in federated multi-task learning (FML), yet existing approaches struggle with heterogeneous environments due to static assumptions or implicit modeling. Moreover, task relationships typically remain implicit, embedded within data distributions and parameter variations, making precise modeling inherently challenging. This challenge is further intensified in Non-IID settings, where data heterogeneity impedes both the identification and accurate estimation of task relationships. To address these challenges, we propose FedSame, a similarity-aware FML framework that leverages Bayesian inference to dynamically model inter-task relationships. The key innovation of FedSame lies in its probabilistic reformulation of task relationship modeling, where an adaptive similarity matrix undergoes continuous Bayesian updates to precisely track evolving task relationships. FedSame’s technical core combines Beta distribution priors with Bayesian update rules, enabling fine-grained detection of subtle variations in task relationship variations during training, and computationally efficient dynamic updates by leveraging the conjugacy property of the Beta distribution. Extensive experiments on two real datasets and a synthetic dataset demonstrate that FedSame consistently outperforms five state-of-the-art baselines in both task relationship modeling accuracy and multi-task classification performance. Remarkably, FedSame attains 73% accuracy in multi-attribute classification on CelebA and 85% accuracy in task relationship modeling on synthetic data, all while maintaining robust performance and notable adaptability in heterogeneous federated environments.

The rapid growth of cloud-based image storage has made privacy-preserving content-based image retrieval increasingly critical for secure and efficient image search. Existing schemes, however, face three key challenges: limited verifiability of search results, inefficient and fault-intolerant support for dynamic updates, and unintended leakage of similarity information. In this paper, we proposeV-FDEIS, a fault-tolerant verifiable dynamic encrypted image search framework for multi-user cloud environments. V-FDEIS simultaneously supports verifiable similarity computation against potentially malicious cloud servers, fault-tolerant dynamic updates in the presence of erroneous client operations, and efficient multi-user search while effectively concealing similarity relationships. A threshold-based reproof mechanism, combined with aggregate verification, ensures low verification overhead even under frequent updates. Extensive theoretical analysis and experimental evaluation demonstrate that V-FDEIS achieves strong privacy guarantees, correctness, fault tolerance, and practical efficiency, outperforming existing approaches.

Federated Learning (FL) has emerged as a pivotal paradigm for multi-client collaborative learning, primarily due to its inherent capability to safeguard privacy. Nonetheless, the heterogeneity among FL clients, characterized by their disparate resource capabilities and not independent and identical (NonIID) local datasets, presents a significant challenge. Specifically, low-resource clients, such as edge devices, grapple with the inadequacy to accommodate the entire model parameter set for training purposes. To mitigate this issue, preceding research has ventured into devising methodologies that entail extracting sub-models from the overarching global model, tailored to the specific communication, computational, and memory constraints of individual clients. Despite these advancements, prevailing sub-model extraction techniques, which predominantly hinge on pre-established rules, overlook a crucial factor: the impact of NonIID local data on the trajectory of neuron update dynamics. This oversight can amplify discrepancies between the practical local updates inferred by the sub-model and those anticipated via the entire model, thereby undermining overall performance. In this paper, we introduceFedGSE, an innovative Gradient-based Neuron Selection methodology designed explicitly for FL environments. This methodology aims to curate sub-models that significantly reduce discrepancies in local updates, enhancing alignment with the global model's learning trajectory. Central to theFedGSEapproach is a sophisticated algorithm that handpicks critical neurons for sub-model construction. These neurons are identified through their pronounced gradient magnitudes, resulting from the training of the global model on a dataset mirroring the client's data distribution. Consequently, the sub-model's induced local gradient updates closely emulate those derived from directly training the client's data on the full global model, fostering enhanced alignment and performance. Extensive experiments over diverse datasets and tasks demonstrate the superiority ofFedGSEover existing baselines.

With the rise of generative paradigms, generative recommendation has garnered increasing attention. The core component is the item code , generally derived by quantizing collaborative or semantic representations to serve as candidate items identifiers in the context. However, existing methods typically construct separate codes for each modality, leading to higher computational and storage costs and hindering the integration of their complementary strengths. Considering this limitation, we seek to integrate two different modalities into a unified code , fully unleashing the potential of complementary nature among modalities. Nevertheless, the integration remains challenging: the integrated embedding obtained by the common concatenation method would lead to underutilization of collaborative knowledge, thereby resulting in limited effectiveness. To address this, we propose a novel method, named UNGER, which integrates semantic and collaborative knowledge into a UNified code for Generative Recommendation . Specifically, we propose to adaptively learn an integrated embedding through the joint optimization of cross-modality knowledge alignment and next-item prediction tasks. Subsequently, to mitigate the information loss caused by the quantization process, we introduce an intra-modality knowledge distillation task, using the integrated embeddings as supervised signals to compensate. Extensive experiments on three widely used benchmarks demonstrate the superiority of our approach compared to existing methods.

The rapid development of diffusion models has greatly advanced AI-generated videos in terms of length and consistency recently, yet assessing AI-generated videos still remains challenging. Previous approaches have often focused on User-Generated Content(UGC), but few have targeted AI-Generated Video Quality Assessment methods. In this work, we introduce MSA-VQA, a Multilevel Semantic-Aware Model for AI-Generated Video Quality Assessment, which leverages CLIP-based semantic supervision and cross-attention mechanisms. Our hierarchical framework analyzes video content at three levels: frame, segment, and video. We propose a Prompt Semantic Supervision Module using text encoder of CLIP to ensure semantic consistency between videos and conditional prompts. Additionally, we propose the Semantic Mutation-aware Module to capture subtle variations between frames. Extensive experiments demonstrate our method achieves state-of-the-art results.

Hallucination remains a significant challenge in Large Vision-Language Models (LVLMs). To alleviate this issue, some methods, known as contrastive decoding, induce hallucinations by manually disturbing the raw vision or instruction inputs and then mitigate them by contrasting the outputs of the original and disturbed LVLMs. However, these holistic input disturbances sometimes induce potential noise and also double the inference cost. To tackle these issues, we propose a simple yet effective method named $\textit{Self-Introspective Decoding}$ (SID). Our empirical investigations reveal that pre-trained LVLMs can introspectively assess the importance of vision tokens based on preceding vision and text (both instruction and generated) tokens. Leveraging this insight, we develop the Context and Text-aware Token Selection (CT$^2$S) strategy, which preserves only the least important vision tokens after the early decoder layers, thereby adaptively amplify vision-and-text association hallucinations during auto-regressive decoding. This strategy ensures that multimodal knowledge absorbed in the early decoder layers induces multimodal contextual rather than aimless hallucinations, and significantly reduces computation burdens. Subsequently, the original token logits subtract the amplified fine-grained hallucinations, effectively alleviating hallucinations without compromising the LVLMs' general ability. Extensive experiments illustrate SID generates less-hallucination and higher-quality texts across various metrics, without much additional computation cost.

Extracting a small subset of crucial rationales from the full input is a key problem in explainability research. The most widely used fundamental criterion for rationale extraction is the maximum mutual information (MMI) criterion. In this paper, we first demonstrate that MMI suffers from diminishing marginal returns. Once part of the rationale has been identified, finding the remaining portions contributes only marginally to increasing the mutual information, making it difficult to use MMI to locate the rest. In contrast to MMI that aims to reproduce the prediction, we seek to identify the parts of the input that the network can actually utilize. This is achieved by comparing how different rationale candidates match the capability space of the weight matrix. The weight matrix of a neural network is typically low-rank, meaning that the linear combinations of its column vectors can only cover part of the directions in a high-dimensional space (high-dimension: the dimensions of an input vector). If an input is fully utilized by the network, it generally matches these directions (e.g., a portion of a hypersphere), resulting in a representation with a high norm. Conversely, if an input primarily falls outside (orthogonal to) these directions, its representation norm will approach zero, behaving like noise that the network cannot effectively utilize. Building on this, we propose using the norms of rationale candidates as an alternative objective to MMI. Through experiments on four text classification datasets and one graph classification dataset using three network architectures (GRUs, BERT, and GCN), we show that our method outperforms MMI and its improved variants in identifying better rationales. We also compare our method with a representative LLM (llama-3.1-8b-instruct) and find that our simple method gets comparable results to it and can sometimes even outperform it.

A major challenge in Federated Learning (FL) is that the data is heterogeneous in terms of features among clients, which significantly degrades the performance of the global model due to its incurred client drift. To solve this problem, some works propose leveraging the global knowledge, e.g., class prototypes, to regularize the output of the local model to mitigate the local model drift. Despite their effectiveness, the input data of the local model are still heterogeneous, causing the parameters to not fully align. Therefore, to fill this gap, this paper proposes a novel method named FedRF which aligns local models on the input side by reusing the intermediate features to approach the global ones. Specifically, we use the intermediate features output by the global model on local data as input to assist in training the local model. Experimental results demonstrate that the proposed method significantly improves the performance of the global model as compared to baselines.

Low-rank adaptation (LoRA) is a widely used parameter-efficient fine-tuning method. In standard LoRA layers, one of the matrices, $A$ or $B$, is initialized to zero, ensuring that fine-tuning starts from the pretrained model. However, there is no theoretical support for this practice. In this paper, we investigate the impact of non-zero initialization on LoRA’s fine-tuning dynamics from an infinite-width perspective. Our analysis reveals that, compared to zero initialization, simultaneously initializing $A$ and $B$ to non-zero values improves LoRA’s robustness to suboptimal learning rates, particularly smaller ones. Further analysis indicates that although the non-zero initialization of $AB$ introduces random noise into the pretrained weight, it generally does not affect fine-tuning performance. In other words, fine-tuning does not need to strictly start from the pretrained model. The validity of our findings is confirmed through extensive experiments across various models and datasets. The code is available at https://github.com/Leopold1423/non_zero_lora-icml25.

To improve the training efficiency of federated learning (FL), previous research has employed low-rank decomposition techniques to reduce communication overhead. In this paper, we seek to enhance the performance of these low-rank decomposition methods. Specifically, we focus on three key issues related to decomposition in FL: what to decompose, how to decompose, and how to aggregate. Subsequently, we introduce three novel techniques: Model Update Decomposition (MUD), Block-wise Kronecker Decomposition (BKD), and Aggregation-Aware Decomposition (AAD), each targeting a specific issue. These techniques are complementary and can be applied simultaneously to achieve optimal performance. Additionally, we provide a rigorous theoretical analysis to ensure the convergence of the proposed MUD. Extensive experimental results show that our approach achieves faster convergence and superior accuracy compared to relevant baseline methods. The code is available at https://github.com/Leopold1423/fedmud-icml25.

Continual Federated Learning (CFL) allows distributed devices to collaboratively learn novel concepts from continuously shifting training data while avoiding \textit{knowledge forgetting} of previously seen tasks. To tackle this challenge, most current CFL approaches rely on extensive rehearsal of previous data. Despite effectiveness, rehearsal comes at a cost to memory, and it may also violate data privacy. Considering these, we seek to apply regularization techniques to CFL by considering their cost-efficient properties that do not require sample caching or rehearsal. Specifically, we first apply traditional regularization techniques to CFL and observe that existing regularization techniques, especially synaptic intelligence, can achieve promising results under homogeneous data distribution but fail when the data is heterogeneous. Based on this observation, we propose a simple yet effective regularization algorithm for CFL named \textbf{FedSSI}, which tailors the synaptic intelligence for the CFL with heterogeneous data settings. FedSSI can not only reduce computational overhead without rehearsal but also address the data heterogeneity issue. Extensive experiments show that FedSSI achieves superior performance compared to state-of-the-art methods.

This study investigates the self-rationalization framework constructed with a cooperative game, where a generator initially extracts the most informative segment from raw input, and a subsequent predictor utilizes the selected subset for its input. The generator and predictor are trained collaboratively to maximize prediction accuracy. In this paper, we first uncover a potential caveat: such a cooperative game could unintentionally introduce a sampling bias during rationale extraction. Specifically, the generator might inadvertently create an incorrect correlation between the selected rationale candidate and the label, even when they are semantically unrelated in the original dataset. Subsequently, we elucidate the origins of this bias using both detailed theoretical analysis and empirical evidence. Our findings suggest a direction for inspecting these correlations through attacks, based on which we further introduce an instruction to prevent the predictor from learning the correlations. Through experiments on six text classification datasets and two graph classification datasets using three network architectures (GRUs, BERT, and GCN), we show that our method significantly outperforms recent rationalization methods.

Semi-supervised Federated Learning (SSFL) is a promising approach that allows clients to collaboratively train a global model in the absence of their local data labels. The key step of SSFL is the re-labeling where each client adopts two types of available models, namely global and local models, to re-label the local data. While various technologies such as using the global model or the average of two models have been proposed to conduct the re-labeling step, little literature delves deeply into the performance dominance and limitations of the two models. In this paper, we first theoretically and empirically demonstrate that the local model achieves higher re-labeling accuracy over local data while the global model can progressively improve the re-labeling performance by introducing the extra data knowledge of other clients. Based on these findings, we propose BSemiFL which re-labels the local data through the collaboration between the local and global model in a Bayesian approach. Specifically, to re-label any given local sample, BSemiFL first uses Bayesian inference to assess the closeness of the local/global model to the sample. Then, it applies a weighted combination of their pseudo labels, using the closeness as the weights. Theoretical analysis shows that the labeling error of our method is smaller than that of simply using the global model, the local model, or their simple average. Experimental results show that BSemiFL improves the performance by up to $9.8\%$ as compared to state-of-the-art methods.

Online Federated Learning (OFL) is a real-time learning paradigm that sequentially executes parameter aggregation immediately for each random arriving client. To motivate clients to participate in OFL, it is crucial to offer appropriate incentives to offset the training resource consumption. However, the design of incentive mechanisms in OFL is constrained by the dynamic variability of Two-sided Incomplete Information (TII) concerning resources, where the server is unaware of the clients’ dynamically changing computational resources, while clients lack knowledge of the real-time communication resources allocated by the server. To incentivize clients to participate in training by offering dynamic rewards to each arriving client, we design a novel Dynamic Bayesian persuasion pricing for online Federated learning (DaringFed) under TII. Specifically, we begin by formulating the interaction between the server and clients as a dynamic signaling and pricing allocation problem within a Bayesian persuasion game, and then demonstrate the existence of a unique Bayesian persuasion Nash equilibrium. By deriving the optimal design of DaringFed under one-sided incomplete information, we further analyze the approximate optimal design of DaringFed with a specific bound under TII. Finally, extensive evaluation conducted on real datasets demonstrate that DaringFed optimizes accuracy and converges speed by 16.99%, while experiments with synthetic datasets validate the convergence of estimate unknown values and the effectiveness of DaringFed in improving the server’s utility by up to 12.6%.

Missing value imputation is a critical data pre-processing technique for machine learning tasks. However, the imputation accuracy would degrade significantly when the volume of known values is limited. To address this issue, this paper proposes a Federated Transformed Sampling Comparison (FTSC) method for imputing missing values, which is based on the optimal transport divergence of two samplings in latent representations space. In detail, assuming that the data samplings come from the same distribution, an invertible neural network (INN) is designed to transform the original sampling data into latent representations within each client. Then, based on the geodesic triangle inequality of the Wasserstein distance, the Wasserstein distance of these latent representations is estimated federatively and modeled as the loss function to jointly optimize the imputation values and the INN parameters. We also theoretically prove the effectiveness of Wasserstein distance estimation and INN training methods. Extensive experiments are implemented on both real and synthetic datasets, and experiment results show that our method in all three missing mechanisms is stable and can outperform several effective imputation methods.

Despite Federated Learning (FL) employing gradient aggregation at the server for distributed training to prevent the privacy leakage of raw data, private information can still be divulged through the analysis of uploaded gradients from clients. Substantial efforts have been made to integrate local differential privacy (LDP) into the system to achieve a strict privacy guarantee. However, existing methods fail to take practical issues into account by merely perturbing each sample with the same mechanism while each client may have their own privacy preferences on privacy-sensitive information (PSI), which is not uniformly distributed across the raw data. In such a case, excessive privacy protection from private-insensitive information can additionally introduce unnecessary noise, which may degrade the model performance. In this work, we study the PSI within data and develop FedRE, that can simultaneously achieve robustness and effectiveness benefits with LDP protection. More specifically, we first define PSI with regard to the privacy preferences of each client. Then, we optimize the LDP by allocating less privacy budget to gradients with higher PSI in a layer-wise manner, thus providing a stricter privacy guarantee for PSI. Furthermore, to mitigate the performance degradation caused by LDP, we design a parameter aggregation mechanism based on the distribution of the perturbed information. We conducted experiments with text tamper detection on T-SROIE and DocTamper datasets, and FedRE achieves competitive performance compared to state-of-the-art methods.

The inference of large language models (LLMs) on distributed edge devices is crucial for privacy-preserving applications. However, its performance is severely degraded in a practically lossy edge network due to frequent synchronization. In this paper, we propose Coda, a novel context-aware distributed acceleration framework tailored for packet loss scenarios. We observe the consistency of sparse patterns of neuron group activations in the same context. By identifying sparsity patterns during the prefill phase and updating neuron groups mapping strategically, we enable a relaxed synchronization for decoding, without slowing down, while preserving accuracy. Our approach offers up to 5.62x speedup, significantly outperforming the state-of-the-art in various scenarios.

Low-rank adaptation (LoRA) is a parameter-efficient fine-tuning (PEFT) method widely used in large language models (LLMs). LoRA essentially describes the projection of an input space into a low-dimensional output space, with the dimensionality determined by the LoRA rank. In standard LoRA, all input tokens share the same weights and undergo an identical input-output projection. This limits LoRA's ability to capture token-specific information due to the inherent semantic differences among tokens. To address this limitation, we propose **Token-wise Projected Low-Rank Adaptation (TopLoRA)**, which dynamically adjusts LoRA weights according to the input token, thereby learning token-wise input-output projections in an end-to-end manner. Formally, the weights of TopLoRA can be expressed as $B\Sigma_X A$, where $A$ and $B$ are low-rank matrices (as in standard LoRA), and $\Sigma_X$ is a diagonal matrix generated from each input token $X$. Notably, TopLoRA does not increase the rank of LoRA weights but achieves more granular adaptation by learning token-wise LoRA weights (i.e., token-wise input-output projections). Extensive experiments across multiple models and datasets demonstrate that TopLoRA consistently outperforms LoRA and its variants. The code is available at https://github.com/Leopold1423/toplora-neurips25.

The current Federated Recommendation System (FedRS) focuses on personalized recommendation services and assumes clients are personalized IoT devices (e.g., Mobile phones). In this paper, we deeply dive into new but practical FedRS applications within the joint venture ecosystem. Subsidiaries engage as participants with their users and items. However, in such a situation, merely exchanging item embedding is insufficient, as user bases always exhibit both overlaps and exclusive segments, demonstrating the complexity of user information. Meanwhile, directly uploading user information is a violation of privacy and unacceptable. To tackle the above challenges, we propose an efficient and privacy-enhanced federated recommendation for the joint venture ecosystem (FR-JVE) that each client transfers more common knowledge from other clients with a distilled user's \textit{rating preference} from the local dataset. More specifically, we first transform the local data into a new format and apply model inversion techniques to distill the rating preference with frozen user gradients before the federated training. Then, a bridge function is employed on each client side to align the local rating preference and aggregated global preference in a privacy-friendly manner. Finally, each client matches similar users to make a better prediction for overlapped users. From a theoretical perspective, we analyze how effectively FR-JVE can guarantee user privacy. Empirically, we show that FR-JVE achieves superior performance compared to state-of-the-art methods.

Federated Continual Learning (FCL) aims to enable sequential privacy-preserving model training on streams of incoming data that vary in edge devices by preserving previous knowledge while adapting to new data. Current FCL literature focuses on restricted data privacy and access to previously seen data while imposing no constraints on the training overhead. This is unreasonable for FCL applications in real-world scenarios, where edge devices are primarily constrained by resources such as storage, computational budget, and label rate. We revisit this problem with a large-scale benchmark and analyze the performance of state-of-the-art FCL approaches under different resource-constrained settings. Various typical FCL techniques and six datasets in two incremental learning scenarios (Class-IL and Domain-IL) are involved in our experiments. Through extensive experiments amounting to a total of over 1,000+ GPU hours, we find that, under limited resource-constrained settings, existing FCL approaches, with no exception, fail to achieve the expected performance. Our conclusions are consistent in the sensitivity analysis. This suggests that most existing FCL methods are particularly too resource-dependent for real-world deployment. Moreover, we study the performance of typical FCL techniques with resource constraints and shed light on future research directions in FCL.

Model-Heterogeneous Federated Learning (Hetero-FL) has attracted growing attention for its ability to aggregate knowledge from heterogeneous models while keeping private data locally. To better aggregate knowledge from clients, ensemble distillation, as a widely used and effective technique, is often employed after global aggregation to enhance the performance of the global model. However, simply combining Hetero-FL and ensemble distillation does not always yield promising results and can make the training process unstable. The reason is that existing methods primarily focus on logit distillation, which, while being model-agnostic with softmax predictions, fails to compensate for the knowledge bias arising from heterogeneous models. To tackle this challenge, we propose a stable and efficient Feature Distillation for model-heterogeneous Federated learning, dubbed FedFD, that can incorporate aligned feature information via orthogonal projection to integrate knowledge from heterogeneous models better. Specifically, a new feature-based ensemble federated knowledge distillation paradigm is proposed. The global model on the server needs to maintain a projection layer for each client-side model architecture to align the features separately. Orthogonal techniques are employed to re-parameterize the projection layer to mitigate knowledge bias from heterogeneous models and thus maximize the distilled knowledge. Extensive experiments show that FedFD achieves superior performance compared to state-of-the-art methods.

Fine-tuning large language models (LLMs) poses significant computational burdens, especially in federated learning (FL) settings. We introduce Layer-wise Efficient Federated Fine-tuning (LEFF), a novel method designed to enhance the efficiency of FL fine-tuning while preserving model performance and minimizing client-side computational overhead. LEFF strategically selects layers for fine-tuning based on client computational capacity, thereby mitigating the straggler effect prevalent in heterogeneous environments. Furthermore, LEFF incorporates an importance-driven layer sampling mechanism, prioritizing layers with greater influence on model performance. Theoretical analysis demonstrates that LEFF achieves a convergence rate of $\mathcal{O}(1/\sqrt{T})$. Extensive experiments on diverse datasets demonstrate that LEFF attains superior computational efficiency and model performance compared to existing federated fine-tuning methods, particularly under heterogeneous conditions.

Multimodal federated learning (MMFL) has gained increasing popularity due to its ability to leverage the correlation between various modalities, meanwhile preserving data privacy for different clients. However, recent studies show that correlation between modalities increase the vulnerability of federated learning against Gradient Inversion Attack (GIA). The complicated situation of MMFL privacy preserving can be summarized as follows: 1) different modality transmits different amounts of information, thus requires various protection strength; 2) correlation between modalities should be taken into account. This paper introduces an information theory perspective to analyze the leaked privacy in process of MMFL, and tries to propose a more reasonable protection method \textbf{Sec-MMFL} based on assessing different information leakage possibilities of each modality by conditional mutual information and adjust the corresponding protection strength. Moreover, we use mutual information to reduce the cross-modality information leakage in MMFL. Experiments have proven that our method can bring more balanced and comprehensive protection at an acceptable cost.

With the increasing sophistication of Large Language Models (LLMs), it is crucial to develop reliable methods to accurately identify whether an interlocutor in real-time dialogue is human or chatbot. However, existing detection methods are primarily designed for analyzing full documents, not the unique dynamics and characteristics of dialogue. These approaches frequently overlook the nuances of interaction that are essential in conversational contexts. This work identifies two key patterns in dialogues: (1) Human-Human (H-H) interactions exhibit significant bidirectional sentiment influence, while (2) Human-Chatbot (H-C) interactions display a clear asymmetric pattern. We propose an innovative approach named ChatbotID, which applies the Granger Causality Test (GCT) to extract a novel set of interactional features that capture the evolving, predictive relationships between conversational attributes. By synergistically fusing these GCT-based interactional features with contextual embeddings, and optimizing the model through a meticulous loss function. Experimental results across multiple datasets and detection models demonstrate the effectiveness of our framework, with significant improvements in accuracy for distinguishing between H-H and H-C dialogues.

Federated Recommendation System (FRS) usually offers recommendation services for users while keeping their data locally to ensure privacy. Currently, most FRS literature assumes that fixed users participate in federated training with personal IoT devices (e.g., mobile phones and PC). However, users may join incrementally, and retraining the entire FRS for each new participating user is unfeasible due to the high training costs and the limited global knowledge contribution from a small number of new users. To guarantee the quality service for these new users, we take a dive into the federated recommendation for cold-start users, a novel scenario where the new participating users can directly obtain a promising recommendation without comprehensive training with all participating users by leveraging both transferred knowledge from the converged warm clients and the knowledge learned from the local data.

Privacy-preserving Cross-Domain Recommendation (CDR) has been extensively studied to address the cold-start problem using auxiliary source domains while simultaneously protecting sensitive information. However, existing privacy-preserving CDR methods rely heavily on transferring sensitive user embeddings or behaviour logs, which leads to adopt privacy methods to distort the data patterns before transferring it to the target domain. The distorted information can compromise overall performance during the knowledge transfer process. To overcome these challenges, our approach differs from existing privacy-preserving methods that focus on safeguarding user-sensitive information. Instead, we concentrate on distilling transferable knowledge from insensitive item embeddings, which we refer to as prototypes. Specifically, we propose a conditional model inversion mechanism to accurately distill prototypes for individual users. We have designed a new data format and corresponding learning paradigm for distilling transferable prototypes from traditional recommendation models using model inversion. These prototypes facilitate bridging the domain shift between distinct source and target domains in a privacy-friendly manner. Additionally, they enable the identification of top-k users in the target domain to substitute for cold-start users prediction. We conduct extensive experiments across large real-world datasets, and the results substantiate the effectiveness of PFCDR https://github.com/walcheng/PFCDR.

Federated learning (FL) has emerged as a significant distributed machine learning paradigm. It allows the training of a global model through user collaboration without the necessity of sharing their original data. Traditional FL generally assumes that each client's data remains fixed or static. However, in real-world scenarios, data typically arrives incrementally, leading to a dynamically expanding data domain. In this study, we examine catastrophic forgetting within Federated Incremental Learning (FIL) and focus on the training resources, where edge clients may not have sufficient storage to keep all data or computational budget to implement complex algorithms designed for the server-based environment. We propose a general and low-cost framework for FIL named Re-Fed+, which is designed to help clients cache important samples for replay. Specifically, when a new task arrives, each client initially caches selected previous samples based on their global and local significance. The client then trains the local model using both the cached samples and the new task samples. From a theoretical perspective, we analyze how effectively Re-Fed+ can identify significant samples for replay to alleviate the catastrophic forgetting issue. Empirically, we show that Re-Fed+ achieves competitive performance compared to state-of-the-art methods.

ABSTRACT Cross‐project defect prediction (CPDP) poses a nontrivial challenge to construct a reliable defect predictor by leveraging data from other projects, particularly when data owners are concerned about data privacy. In recent years, federated learning (FL) has become an emerging paradigm to guarantee privacy information by collaborative training a global model among multiple parties without sharing raw data. While the direct application of FL to the CPDP task offers a promising solution to address privacy concerns, the data heterogeneity arising from proprietary projects across different companies or organizations will bring troubles for model training. In this paper, we study the privacy‐preserving CPDP with data heterogeneity under the FL framework. To address this problem, we propose a novel knowledge enhancement approach named FedDP with two simple but effective solutions: 1. local heterogeneity awareness and 2. global knowledge distillation. Specifically, we employ open‐source project data as the distillation dataset and optimize the global model with the heterogeneity‐aware local model ensemble via knowledge distillation. Experimental results on 19 projects from two datasets demonstrate that our method significantly outperforms baselines.

Abstract Rationalization is a framework that aims to build self-explanatory NLP models by extracting a subset of human-intelligible pieces of their inputting texts. It involves a cooperative game where a selector selects the most human-intelligible parts of the input as the rationale, followed by a predictor that makes predictions based on these selected rationales. Existing literature uses the cross-entropy between the model’s predictions and the ground-truth labels to measure the informativeness of the selected rationales, guiding the selector to choose better ones. In this study, we first theoretically analyze the objective of rationalization by decomposing it into two parts: the model-agnostic informativeness of the rationale candidates and the predictor’s degree of fit. We then provide various empirical evidence to support that, under this framework, the selector tends to sample from a limited small region, causing the predictor to overfit these localized areas. This results in a significant mismatch between the cross-entropy objective and the informativeness of the rationale candidates, leading to suboptimal solutions. To address this issue, we propose a simple yet effective method that introduces random vicinal1 perturbations to the selected rationale candidates. This approach broadens the predictor’s assessment to a vicinity around the selected rationale candidate. Compared to recent competitive methods, our method significantly improves rationale quality (by up to 6.6%) across six widely used classification datasets. The term “vicinal” is borrowed from vicinal risk minimization (Chapelle et al., 2000); “vicinal” means neighboring or adjacent.

Inferring user status at the edge is essential for delivering personalized services, such as detecting emotional states. However, deploying large-scale models directly on user devices is impractical due to substantial computational overhead and the scarcity of labeled data. Conversely, uploading raw data to the cloud for processing raises significant privacy concerns and incurs prohibitive communication costs. To address this challenge, we propose a privacy-preserving multimodal inference framework that leverages large-scale public data while safeguarding sensitive information and optimizing computational efficiency. Specifically, we first train a teacher model in the cloud using publicly available data. Through a feature distillation process, the knowledge from this teacher model is transferred to a lightweight encoder deployed at the user end. This transfer is tailored to the user's data, ensuring that only relevant knowledge is distilled. To accommodate varying communication constraints, we introduce a feature compression mechanism that significantly reduces communication overhead without compromising inference accuracy. Extensive experiments on emotion recognition tasks demonstrate that the proposed framework effectively balances privacy preservation, resource efficiency, and inference accuracy, facilitating seamless collaboration between cloud and edge devices.

Personalized Federated Learning (pFL) can improve the accuracy of local models and provide enhanced edge intelligence without exposing the raw data in Mobile Edge Computing (MEC). However, in the MEC environment with constrained communication resources, transmitting the entire model between the server and the clients in traditional pFL methods imposes substantial communication overhead, which can lead to inaccurate personalization and degraded performance of mobile clients. In response, we propose a Communication-Efficient pFL architecture to enhance the performance of personalized models while minimizing communication overhead in MEC. First, a Knowledge-Aware Parameter Coaching method (KAPC) is presented to produce a more accurate personalized model by utilizing the layer-wise parameters of other clients with adaptive aggregation weights. Then, convergence analysis of the proposed KAPC is developed in both the convex and non-convex settings. Second, a Bidirectional Layer Selection algorithm (BLS) based on self-relationship and generalization error is proposed to select the most informative layers for transmission, which reduces communication costs. Extensive experiments are conducted, and the results demonstrate that the proposed KAPC achieves superior accuracy compared to the state-of-the-art baselines, while the proposed BLS substantially improves resource utilization without sacrificing performance.

Federated Learning (FL) is to collaboratively train a global model among distributed clients by iteratively aggregating their local updates without sharing their raw data, whereby the global modal can approximately converge to the centralized training way over a global dataset that composed of all local datasets (i.e., union of all users’ local data). However, in real-world scenarios, the distributions of the data classes are often imbalanced not only locally, but also in the global dataset, which severely deteriorate the FL performance due to the conflicting knowledge aggregation. Existing solutions for FL class imbalance either focus on the local data to regulate the training process or purely aim at the global datasets, which often fail to alleviate the class imbalance problem if there is mismatch between the local and global imbalance. Considering these limitations, this paper proposes a Global-Local Joint Learning method, namely GLJL, which simultaneously harmonizes the global and local class imbalance issue by jointly embedding the local and the global factors into each client’s loss function. Through extensive experiments over popular datasets with various class imbalance settings, we show that the proposed method can significantly improve the model accuracy over minority classes without sacrificing the accuracy of other classes.

Federated learning (FL) is an emerging distributed machine learning paradigm that enables collaborative model training without sharing local data. Despite its advantages, FL suffers from substantial communication overhead, which can affect training efficiency. Recent efforts have mitigated this issue by quantizing model updates to reduce communication costs. However, most existing methods apply quantization only after local training, introducing quantization errors into the trained parameters and potentially degrading model accuracy. In this paper, we propose Federated Bit Freezing (FedBiF), a novel FL framework that directly learns quantized model parameters during local training. In each communication round, the server first quantizes the model parameters and transmits them to the clients. FedBiF then allows each client to update only a single bit of the multi-bit parameter representation, freezing the remaining bits. This bit-by-bit update strategy reduces each parameter update to one bit while maintaining high precision in parameter representation. Extensive experiments are conducted on five widely used datasets under both IID and Non-IID settings. The results demonstrate that FedBiF not only achieves superior communication compression but also promotes sparsity in the resulting models. Notably, FedBiF attains accuracy comparable to FedAvg, even when using only 1 bit-per-parameter (bpp) for uplink and 3 bpp for downlink communication. The code is available athttps://github.com/Leopold1423/fedbif-tpds25.

Few-shot intent classification and slot filling are important but challenging tasks due to the scarcity of finely labeled data. Therefore, current works first train a model on source domains with sufficiently labeled data, and then transfer the model to target domains where only rarely labeled data is available. However, experience transferring as a whole usually suffers from gaps that exist among source domains and target domains. For instance, transferring domain-specific-knowledge-related experience is difficult. To tackle this problem, we propose a new method that explicitly decouples the transferring of general-semantic-representation-related experience and the domain-specific-knowledge-related experience. Specifically, for domain-specific-knowledge-related experience, we design two modules to capture intent-slot relation and slot-slot relation respectively. Extensive experiments on Snips and FewJoint datasets show that our method achieves state-of-the-art performance. The method improves the joint accuracy metric from 27.72% to 42.20% in the 1-shot setting, and from 46.54% to 60.79% in the 5-shot setting.

Open-World Compositional Zero-shot Learning (OW-CZSL) aims to recognize novel compositions of state and object primitives in images with no priors on the compositional space, which induces a tremendously large output space containing all possible state-object compositions. Existing works either learn the joint compositional state-object embedding or predict simple primitives with separate classifiers. However, the former method heavily relies on external word embedding methods, and the latter ignores the interactions of interdependent primitives, respectively. In this paper, we revisit the primitive prediction approach and propose a novel method, termed Progressive Cross-primitive Compatibility (ProCC), to mimic the human learning process for OW-CZSL tasks. Specifically, the cross-primitive compatibility module explicitly learns to model the interactions of state and object features with the trainable memory units, which efficiently acquires cross-primitive visual attention to reason high-feasibility compositions, without the aid of external knowledge. Moreover, to alleviate the invalid cross-primitive interactions, especially for partial-supervision conditions (pCZSL), we design a progressive training paradigm to optimize the primitive classifiers conditioned on pre-trained features in an easy-to-hard manner. Extensive experiments on three widely used benchmark datasets demonstrate that our method outperforms other representative methods on both OW-CZSL and pCZSL settings by large margins.

This paper investigates a new, practical, but challenging problem named Non-exemplar Online Class-incremental continual Learning (NO-CL), which aims to preserve the discernibility of base classes without buffering data examples and efficiently learn novel classes continuously in a single-pass (i.e., online) data stream. The challenges of this task are mainly two-fold: (1) Both base and novel classes suffer from severe catastrophic forgetting as no previous samples are available for replay. (2) As the online data can only be observed once, there is no way to fully re-train the whole model, e.g., re-calibrate the decision boundaries via prototype alignment or feature distillation. In this paper, we propose a novel Dual-prototype Self-augment and Refinement method (DSR) for NO-CL problem, which consists of two strategies: 1) Dual class prototypes: vanilla and high-dimensional prototypes are exploited to utilize the pre-trained information and obtain robust quasi-orthogonal representations rather than example buffers for both privacy preservation and memory reduction. 2) Self-augment and refinement: Instead of updating the whole network, we optimize high-dimensional prototypes alternatively with the extra projection module based on self-augment vanilla prototypes, through a bi-level optimization problem. Extensive experiments demonstrate the effectiveness and superiority of the proposed DSR in NO-CL.

Personalized Federated Learning (pFL) can effectively exploit the non-IID data from distributed clients by customizing personalized models. Existing pFL methods either simply take the local model as a whole for aggregation or require significant training overhead to induce the inter-client personalized weights, and thus clients cannot efficiently exploit the mutually relevant knowledge from each other. In this paper, we propose a knowledge-aware parameter coaching scheme where each client can swiftly and granularly refer to parameters of other clients to guide the local training, whereby accurate personalized client models can be efficiently produced without contradictory knowledge. Specifically, a novel regularizer is designed to conduct layer-wise parameters coaching via a relation cube, which is constructed based on the knowledge represented by the layered parameters among all clients. Then, we develop an optimization method to update the relation cube and the parameters of each client. It is theoretically demonstrated that the convergence of the proposed method can be guaranteed under both convex and non-convex settings. Extensive experiments are conducted over various datasets, which show that the proposed method can achieve better performance compared with the state-of-the-art baselines in terms of accuracy and convergence speed.

Modern Autonomous Vehicles (AVs) implement the Visual Perception Module (VPM) to perceive their surroundings. This VPM adopts various Deep Neural Network (DNN) models to process the data collected from cameras and LiDAR. Prior studies have shown that these models are vulnerable to physical adversarial examples (PAEs), which pose a critical safety risk to the autonomous driving task. While a few defense methods have been proposed to safeguard AVs, most of them only target a limited set of attack types and specific scenarios, making them impractical for real-world protection.

In Federated Learning (FL), the data in each client is typically assumed fixed or static. However, data often comes in an incremental manner in real-world applications, where the data domain may increase dynamically. In this work, we study catastrophic forgetting with data heterogeneity in Federated Incremental Learning (FIL) scenarios where edge clients may lack enough storage space to retain full data. We propose to employ a simple, generic frame-work for FIL named Re-Fed, which can coordinate each client to cache important samples for replay. More specifically, when a new task arrives, each client first caches selected previous samples based on their global and local im-portance. Then, the client trains the local model with both the cached samples and the samples from the new task. The-oretically, we analyze the ability of Re-Fed to discover important samples for replay thus alleviating the catastrophic forgetting problem. Moreover, we empirically show that Re-Fed achieves competitive performance compared to state-of-the-art methods.

Existing Knowledge Distillation (KD) methods typically focus on transferring knowledge from a large-capacity teacher to a low-capacity student model, achieving sub-stantial success in unimodal knowledge transfer. However, existing methods can hardly be extended to Cross-Modal Knowledge Distillation (CMKD), where the knowledge is transferred from a teacher modality to a different student modality, with inference only on the distilled student modality. We empirically reveal that the modality gap, i.e., modality imbalance and soft label misalignment, incurs the in-effectiveness of traditional KD in CMKD. As a solution, we propose a novel f_ustomized Crossmodal Knowledge Distillation (C2KD). Specifically, to alleviate the modality gap, the pre-trained teacher performs bidirectional distillation with the student to provide customized knowledge. The On-the-Fly Selection Distillationi OFSD) strategy is applied to selectively filter out the samples with misaligned soft labels, where we distill cross-modal knowledge from non-target classes to avoid the modality imbalance issue. To further provide receptive cross-modal knowledge, proxy student and teacher, inheriting unimodal and cross-modal knowledge, is formulated to progressively transfer cross-modal knowledge through bidirectional distillation. Experimental results on audio-visual, image-text, and RGB-depth datasets demonstrate that our method can effectively transfer knowledge across modalities, achieving superior performance against traditional KD by a large margin.

Rationalization models, which select a subset of input text as rationale—crucial for humans to understand and trust predictions—have recently emerged as a prominent research area in eXplainable Artificial Intelligence (XAI). However, most of previous studies mainly focus on improving the quality of the rationale, ignoring its robustness to malicious attack. Specifically, whether the rationalization models can still generate high-quality rationale under the adversarial attack remains unknown. To explore this, this paper proposes UAT2E, which aims to undermine the explainability of rationalization models without altering their predictions, thereby eliciting distrust in these models from human users. UAT2E employs the gradient-based search on triggers and then inserts them into the original input to conduct both the non-target and target attack. Experimental results on five datasets reveal the vulnerability of rationalization models in terms of explanation, where they tend to select more meaningless tokens under attacks. Based on this, we make a series of recommendations for improving rationalization models in terms of explanation.

Deep learning models, usually trained on large datasets, have witnessed great achievements for both classification and regression tasks during past years. However, in practical settings, the datasets are commonly imbalanced where the number of data samples differs among labels (i.e.,categories or target values), leading to serious performance degradation of the trained models. Although many prior works have been proposed to solve the data imbalance problem for the classification task, there are still limited works considering the regression task. In this work, we identify the distinct challenges in regression problems compared to traditional classification problems, such as managing the continuous nature of the target variable and navigating fuzzy decision-making boundaries. To address these challenges, we propose an Adaptive Weighted Statistical Space Smoothing method (AWSSS), which alleviates the impact of imbalanced data by learning from continuously valued, imbalanced data and extending local similarity across the entire target range. AWSSS extracts statistical values from the original dataset and smoothes statistical features. Subsequently, it aligns features across different regions and applies adaptive weights to facilitate knowledge transfer between adjacent areas. Finally, AWSSS calculates the loss values, which are used to refine the model’s parameters through iterative updates. Extensive experiments conducted on various datasets demonstrate the effectiveness of the proposed method as compared to state-of-the-art methods.

Federated Learning (FL) has emerged as a foundational paradigm that enables collaborative training of deep neural networks across distributed clients while ensuring data privacy. However, most of the existing research primarily concentrates on supervised federated learning tailored for specific downstream tasks. In this paper, we concentrate on unsupervised federated learning, where clients collaborate to identify data patterns, structures, or representations. Contrastive learning methods have proven highly effective for unsupervised learning in server-based environments. However, a straightforward adaptation of the contrastive learning approach for this setting falls short. Upon analyzing client behaviors during training, we identified an unstable training process stemming from NonIID issues, which can result in diminished performance. To address this problem, we propose a novel Temperature-Adaption based unsupervised Federated prototype learning approach, termed FedTA. This method incorporates two straightforward yet potent solutions: 1. Prototype Enhancement and 2. Temperature-Adaption. Experimental results show that our proposed approach surpasses the state-of-the-art, achieving a performance accuracy improvement of up to 5.3%.

Estimating causal effects among different events is of great importance to critical fields such as drug development. Nevertheless, the data features associated with events may be distributed across various silos and remain private within respective parties, impeding direct information exchange between them. This, in turn, can result in biased estimations of local causal effects, which rely on the characteristics of only a subset of the covariates. To tackle this challenge, we introduce an innovative disentangle architecture designed to facilitate the seamless cross-silo transmission of model parameters, enriched with causal mechanisms, through a combination of shared and private branches. Besides, we introduce global constraints into the equation to effectively mitigate bias within the various missing domains, thereby elevating the accuracy of our causal effect estimation. Extensive experiments conducted on new semi-synthetic datasets show that our method outperforms state-of-the-art baselines.

Semantic communication systems (SCS) have gained extensive attention with the advancement of Artificial Intelligence (AI), which transmits the data feature instead of the raw bits, whereby the communication efficiency can be substantially enhanced, e.g., via a neural network or encoder to convert from the massive user data to corresponding light-weight feature map. However, SCS can be vulnerable to adversarial noise when transmitting the feature data, which may mislead the downstream tasks at the receiver side, e.g., leading to misclassification due to the disturbed receiving information. In this paper, we investigate the overshadowing-based attacks by perturbing the physical signal with artificial adversarial noise during the semantic feature transmission. Specifically, we directly attack the waveform after the modulation of the feature bits, and conduct both the white-box and black-box attacks to evaluate the vulnerability. In our attack methods, we use the local transfer model to acquire the gradient details and provide the gradient-based strategy for generating the perturbation. The experiment results demonstrate that both white-box and black-box attacks can be a critical threat for SCS and significantly degrade the performance of downstream tasks.

Rationalization empowers deep learning models with self-explaining capabilities through a cooperative game, where a generator selects a semantically consistent subset of the input as a rationale, and a subsequent predictor makes predictions based on the selected rationale. In this paper, we discover that rationalization is prone to a problem named rationale shift, which arises from the algorithmic bias of the cooperative game. Rationale shift refers to a situation where the semantics of the selected rationale may deviate from the original input, but the predictor still produces accurate predictions based on the deviation, resulting in a compromised generator with misleading feedback. To address this issue, we first demonstrate the importance of the alignment between the rationale and the full input through both empirical observations and theoretical analysis. Subsequently, we introduce a novel approach called DAR (Discriminatively Aligned Rationalization), which utilizes an auxiliary module pretrained on the full input to discriminatively align the selected rationale and the original input. We theoretically illustrate how DAR accomplishes the desired alignment, thereby overcoming the rationale shift problem. The experiments on two widely used real-world benchmarks show that the proposed method significantly improves the explanation quality (measured by the overlap between the model-selected explanation and the human-annotated rationale) as compared to state-of-the-art techniques. Additionally, results on two synthetic settings further validate the effectiveness of DAR in addressing the rationale shift problem.

Distributed full-graph training of Graph Neural Networks (GNNs) has been widely adopted to learn large-scale graphs. While recent system advancements can improve the training throughput of GNNs, their practical adoption is limited by the potential accuracy decline. This concern is particularly prominent in deeper and more intricate GNN architectures, where noticeable performance degradation becomes apparent. Moreover, existing works fail to comprehensively consider diverse opportunities for acceleration. Motivated by these deficiencies, we propose Sylvie,a full-graph training system that not only improves the training throughput substantially but also maintains the model quality for universal GNNs. By harnessing the inherent information embedded in the graph data and model structure, Sylvie intelligently optimizes GNN training across three key dimensions: data, time, and execution. It identifies performance-relevant features of the input graph offline as subsequent optimization guidance. Subsequently, Sylvie devises an online convergence-maintenance strategy that adaptively integrates and aligns GNN-specific quantization and inter-epoch asynchronous training with the real-time training characteristics. Extensive experiments demonstrate that Sylvie surpasses existing GNN training systems by up to 17.2× speedup for both shallow and deep GNNs, without compromising the model accuracy.

In Federated Learning (FL), model aggregation is pivotal. It involves a global server iteratively aggregating client local trained models in successive rounds without accessing private data. Traditional methods typically aggregate the local models from the current round alone. However, due to the statistical heterogeneity across clients, the local models from different clients may be greatly diverse, making the obtained global model incapable of maintaining the specific knowledge of each local model. In this paper, we introduce a novel method, FedCDA, which selectively aggregates cross-round local models, decreasing discrepancies between the global model and local models. The principle behind FedCDA is that due to the different global model parameters received in different rounds and the non-convexity of deep neural networks, the local models from each client may converge to different local optima across rounds. Therefore, for each client, we select a local model from its several recent local models obtained in multiple rounds, where the local model is selected by minimizing its divergence from the local models of other clients. This ensures the aggregated global model remains close to all selected local models to maintain their data knowledge. Extensive experiments conducted on various models and datasets reveal our approach outperforms state-of-the-art aggregation methods.

Federated learning is a promising distributed machine learning paradigm that can effectively exploit large-scale data without exposing users’ privacy. However, it may incur significant communication overhead, thereby potentially impairing the training efficiency. To address this challenge, numerous studies suggest binarizing the model updates. Nonetheless, traditional methods usually binarize model updates in a post-training manner, resulting in significant approximation errors and consequent degradation in model accuracy. To this end, we propose Federated Binarization-Aware Training (FedBAT), a novel framework that directly learns binary model updates during the local training process, thus inherently reducing the approximation errors. FedBAT incorporates an innovative binarization operator, along with meticulously designed derivatives to facilitate efficient learning. In addition, we establish theoretical guarantees regarding the convergence of FedBAT. Extensive experiments are conducted on four popular datasets. The results show that FedBAT significantly accelerates the convergence and exceeds the accuracy of baselines by up to 9%, even surpassing that of FedAvg in some cases.

Transformer model empowered architectures have become a pillar of cloud services that keeps reshaping our society. However, the dynamic query loads and heterogeneous user requirements severely challenge current transformer serving systems, which rely on pre-training multiple variants of a foundation model, i.e., with different sizes, to accommodate varying service demands. Unfortunately, such a mechanism is unsuitable for large transformer models due to the additional training costs and excessive I/O delay. In this paper, we introduce OTAS, the first elastic serving system specially tailored for transformer models by exploring lightweight token management. We develop a novel idea called token adaptation that adds prompting tokens to improve accuracy and removes redundant tokens to accelerate inference. To cope with fluctuating query loads and diverse user requests, we enhance OTAS with application-aware selective batching and online token adaptation. OTAS first batches incoming queries with similar service-level objectives to improve the ingress throughput. Then, to strike a tradeoff between the overhead of token increment and the potentials for accuracy improvement, OTAS adaptively adjusts the token execution strategy by solving an optimization problem. We implement and evaluate a prototype of OTAS with multiple datasets, which show that OTAS improves the system utility by at least 18.2%.

Federated Learning (FL) suffers from severe performance degradation due to the data heterogeneity among clients. Some existing work suggests that the fundamental reason is that data heterogeneity can cause local model drift, and therefore proposes to calibrate the direction of local updates to solve this problem. Though effective, existing methods generally take the model as a whole, which lacks a deep understanding of how the neurons within deep classification models evolve during local training to form model drift. In this paper, we bridge this gap by performing an intuitive and theoretical analysis of the activation changes of each neuron during local training. Our analysis shows that the high activation of some neurons on the samples of a certain class will be reduced during local training when these samples are not included in the client, which we call neuron drift, thus leading to the performance reduction of this class. Motivated by this, we propose a novel and simple algorithm called FedNLR, which utilizes Neuron-wise Learning Rates during the FL local training process. The principle behind this is to enhance the learning of neurons bound to local classes on local data knowledge while reducing the decay of non-local classes knowledge stored in neurons. Experimental results demonstrate that FedNLR achieves state-of-the-art performance on federated learning with popular deep neural networks.

Recently, Multimodal Learning (MML) has gained significant interest as it compensates for single-modality limitations through comprehensive complementary information within multimodal data. However, traditional MML methods generally use the joint learning framework with a uniform learning objective that can lead to the modality competition issue, where feedback predominantly comes from certain modalities, limiting the full potential of others. In response to this challenge, this paper introduces DI-MML, a novel detached MML framework designed to learn complementary information across modalities under the premise of avoiding modality competition. Specifically, DI-MML addresses competition by separately training each modality encoder with isolated learning objectives. It further encourages cross-modal interaction via a shared classifier that defines a common feature space and employing a dimension-decoupled unidirectional contrastive (DUC) loss to facilitate modality-level knowledge transfer. Additionally, to account for varying reliability in sample pairs, we devise a certainty-aware logit weighting strategy to effectively leverage complementary information at the instance level during inference. Extensive experiments conducted on audio-visual, flow-image, and front-rear view datasets show the superior performance of our proposed method. The code is released at https://github.com/fanyunfeng-bit/DI-MML.

Federated learning is a promising distributed training paradigm that effectively safeguards data privacy. However, it may involve significant communication costs, which hinders training efficiency. In this paper, we aim to enhance communication efficiency from a new perspective. Specifically, we request the distributed clients to find optimal model updates relative to global model parameters within predefined random noise. For this purpose, we propose Federated Masked Random Noise (FedMRN), a novel framework that enables clients to learn a 1-bit mask for each model parameter and apply masked random noise (i.e., the Hadamard product of random noise and masks) to represent model updates. To make FedMRN feasible, we propose an advanced mask training strategy, called progressive stochastic masking (PSM). After local training, each client only need to transmit local masks and a random seed to the server. Additionally, we provide theoretical guarantees for the convergence of FedMRN under both strongly convex and non-convex assumptions. Extensive experiments are conducted on four popular datasets. The results show that FedMRN exhibits superior convergence speed and test accuracy compared to relevant baselines, while attaining a similar level of accuracy as FedAvg.

Multi-modal domain generalization (MMDG) requires that models trained on multi-modal source domains can generalize to unseen target distributions with the same modality set. Sharpness-aware minimization (SAM) is an effective technique for traditional uni-modal domain generalization (DG), however, with limited improvement in MMDG. In this paper, we identify that modality competition and discrepant uni-modal flatness are two main factors that restrict multi-modal generalization. To overcome these challenges, we propose to construct consistent flat loss regions and enhance knowledge exploitation for each modality via cross-modal knowledge transfer. Firstly, we turn to the optimization on representation-space loss landscapes instead of traditional parameter space, which allows us to build connections between modalities directly. Then, we introduce a novel method to flatten the high-loss region between minima from different modalities by interpolating mixed multi-modal representations. We implement this method by distilling and optimizing generalizable interpolated representations and assigning distinct weights for each modality considering their divergent generalization capabilities. Extensive experiments are performed on two benchmark datasets, EPIC-Kitchens and Human-Animal-Cartoon (HAC), with various modality combinations, demonstrating the effectiveness of our method under multi-source and single-source settings. Our code is open-sourced.

An important line of research in the field of explainability is to extract a small subset of crucial rationales from the full input. The most widely used criterion for rationale extraction is the maximum mutual information (MMI) criterion. However, in certain datasets, there are spurious features non-causally correlated with the label and also get high mutual information, complicating the loss landscape of MMI. Although some penalty-based methods have been developed to penalize the spurious features (e.g., invariance penalty, intervention penalty, etc) to help MMI work better, these are merely remedial measures. In the optimization objectives of these methods, spurious features are still distinguished from plain noise, which hinders the discovery of causal rationales. This paper aims to develop a new criterion that treats spurious features as plain noise, allowing the model to work on datasets rich in spurious features as if it were working on clean datasets, thereby making rationale extraction easier. We theoretically observe that removing either plain noise or spurious features from the input does not alter the conditional distribution of the remaining components relative to the task label. However, significant changes in the conditional distribution occur only when causal features are eliminated. Based on this discovery, the paper proposes a criterion for \textbf{M}aximizing the \textbf{R}emaining \textbf{D}iscrepancy (MRD). Experiments on six widely used datasets show that our MRD criterion improves rationale quality (measured by the overlap with human-annotated rationales) by up to $10.4\%$ as compared to several recent competitive MMI variants. Code: \url{https://github.com/jugechengzi/Rationalization-MRD}.

Sub-model extraction based federated learning has emerged as a popular strategy for training models on resource-constrained devices. However, existing methods treat all clients equally and extract sub-models using predetermined rules, which disregard the statistical heterogeneity across clients and may lead to fierce competition among them. Specifically, this paper identifies that when making predictions, different clients tend to activate different neurons of the entire model related to their respective distributions. If highly activated neurons from some clients with one distribution are incorporated into the sub-model allocated to other clients with different distributions, they will be forced to fit the new distributions, which can hinder their activation over the previous clients and result in a performance reduction. Motivated by this finding, we propose a novel method called FedDSE, which can reduce the conflicts among clients by extracting sub-models based on the data distribution of each client. The core idea of FedDSE is to empower each client to adaptively extract neurons from the entire model based on their activation over the local dataset. We theoretically show that FedDSE can achieve an improved classification score and convergence over general neural networks with the ReLU activation function. Experimental results on various datasets and models show that FedDSE outperforms all state-of-the-art baselines.

The escalating proliferation of user generated contents such as videos and images are dominating the network traffic. The optimal strategy for mitigating backbone congestion and minimizing user request latency lies in prudent caching at edge stations within distributed autonomous networks, obviating the necessity to transmit data to the cloud. However, accurately caching content based on distributed autonomous networks requires elaborative collaboration between edge servers, which remains a great challenge, especially when content is highly dynamic and the storage resources of edge stations are limited. To tackle this challenge, this paper proposes a contextual bandit-based online caching algorithm for evaluating the optimal content hit rate reward, which can adapt to the constantly changing stream of emerging content. We build the content space, BS space, and a fine-grained space searching method to cache contents and corresponding edge stations. Furthermore, to perform collaborative caching and sharing between edges, we propose a federated autonomous multi-layer caching framework, whereby each server can locally learn the model for accurate caching and a synchronous mechanism is set up for global updating, further improving the hit rates. Finally, we perform theoretical proofs and simulations, demonstrating that our regret is sublinear and our caching algorithm outperforms several state-of-the-art algorithms.

Machine learning models are widely applied for vehicular applications, which are essential to future intelligent transportation system (ITS). Traditional model training methods commonly employ a client-server architecture to perform local training and global iterative aggregations, which can consume significant bandwidth resources that are often absent in vehicular networks, especially in high vehicle density scenarios. Modern vehicle users naturally can collaboratively train machine learning models as they are the data owner and have strong local computing power from the onboard units (OBU). In this paper, we propose a novel collaborative learning scheme for mobile vehicles that can utilize the opportunistic vehicle-to-roadside (V2R) communication to exploit the common priors of vehicular data without interaction with a centralized coordinator. Specifically, vehicles perform local training during the driving journey, and simply upload its local model to roadside unit (RSU) encountered on the way. RSU's model will be updated accordingly and sent back to the vehicle via the V2R communication. We have theoretically shown that RSUs' models can eventually converge without a backhaul connection. Extensive experiments upon various road configurations demonstrate that the proposed scheme can efficiently train models among vehicles without dedicated Internet access and scale well with both the road range and vehicle density.

Packet loss inferring can enable a transceiver to distinguish between channel impairment and collision for transmission failures, and thus can improve the network performance by exclusively performing rate adaptation or adjusting the medium access parameter. Machine learning methods from literature have shown great potential in producing models that can detect the loss causes over various network trace, however haven't considered accurate data-driven loss inferring on resource-constrained devices that cannot accommodate deep models. In this paper, we propose a novel packet loss inferring framework that can train lightweight models to distinguish between channel losses and collisions by learning the data trace from both simulation and real devices. Specifically, we first train a sophisticated teacher model based on extensive simulation datasets, whose knowledge is then transferred to a small student model that can be deployed on tiny device. The simulation-reality distillation is conducted via personalized trace from each client correspondingly, whose performance bound is analytically guaranteed. We have implemented our method on real testbed and show that the network access performance can be significantly improved, especially for sudden network variations.

This paper addresses the temporal sentence grounding (TSG). Although existing methods have made decent achievements in this task, they not only severely rely on abundant video-query paired data for training, but also easily fail into the dataset distribution bias. To alleviate these limitations, we introduce a novel Equivariant Consistency Regulation Learning (ECRL) framework to learn more discriminative query-related frame-wise representations for each video, in a self-supervised manner. Our motivation comes from that the temporal boundary of the query-guided activity should be consistently predicted under various video-level transformations. Concretely, we first design a series of spatio-temporal augmentations on both foreground and background video segments to generate a set of synthetic video samples. In particular, we devise a self-refine module to enhance the completeness and smoothness of the augmented video. Then, we present a novel self-supervised consistency loss (SSCL) applied on the original and augmented videos to capture their invariant query-related semantic by minimizing the KL-divergence between the sequence similarity of two videos and a prior Gaussian distribution of timestamp distance. At last, a shared grounding head is introduced to predict the transform-equivariant query-guided segment boundaries for both the original and augmented videos. Extensive experiments on three challenging datasets (ActivityNet, TACoS, and Charades-STA) demonstrate both effectiveness and efficiency of our proposed ECRL framework.

Federated Learning (FL) is to allow multiple clients to collaboratively train a model while keeping their data locally. However, existing FL approaches typically assume that the data in each client is static and fixed, which cannot account for incremental data with domain shift, leading to catastrophic forgetting on previous domains, particularly when clients are common edge devices that may lack enough storage to retain full samples of each domain. To tackle this challenge, we proposeFederatedDomain-IncrementalLearning viaSynergisticReplay (SR-FDIL), which alleviates catastrophic forgetting by coordinating all clients to cache samples and replay them. More specifically, when new data arrives, each client selects the cached samples based not only on their importance in the local dataset but also on their correlation with the global dataset. Moreover, to achieve a balance between learning new data and memorizing old data, we propose a novel client selection mechanism by jointly considering the importance of both old and new data. We conducted extensive experiments on several datasets of which the results demonstrate that SR-FDIL outperforms state-of-the-art methods by up to 4.05% in terms of average accuracy of all domains.

Rationalization is to employ a generator and a predictor to construct a self-explaining NLP model in which the generator selects a subset of human-intelligible pieces of the input text to the following predictor.However, rationalization suffers from two key challenges, i.e., spurious correlation and degeneration, where the predictor overfits the spurious or meaningless pieces solely selected by the not-yet well-trained generator and in turn deteriorates the generator.Although many studies have been proposed to address the two challenges, they are usually designed separately and do not take both of them into account.In this paper, we propose a simple yet effective method named MGR to simultaneously solve the two problems.The key idea of MGR is to employ multiple generators such that the occurrence stability of real pieces is improved and more meaningful pieces are delivered to the predictor.Empirically 1 , we show that MGR improves the F1 score by up to 20.9% as compared to state-of-the-art methods.

Multimodal learning (MML) aims to jointly exploit the common priors of different modalities to compensate for their inherent limitations. However, existing MML methods often optimize a uniform objective for different modalities, leading to the notorious “modality imbalance” problem and counterproductive MML performance. To address the problem, some existing methods modulate the learning pace based on the fused modality, which is dominated by the better modality and eventually results in a limited improvement on the worse modal. To better exploit the features of multimodal, we propose Prototypical Modality Rebalance (PMR) to perform stimulation on the particular slow-learning modality without interference from other modalities. Specifically, we introduce the prototypes that represent general features for each class, to build the non-parametric classifiers for uni-modal performance evaluation. Then, we try to accelerate the slow-learning modality by enhancing its clustering toward prototypes. Furthermore, to alleviate the suppression from the dominant modality, we introduce a prototype-based entropy regularization term during the early training stage to prevent premature convergence. Besides, our method only relies on the representations of each modality and without restrictions from model structures and fusion methods, making it with great application potential for various scenarios. The source code is available here11https://github.com/fanyunfeng-bit/Modal-Imbalance-PMR.

Federated Distillation (FD) has recently attracted increasing attention for its efficiency in aggregating multiple diverse local models trained from statistically heterogeneous data of distributed clients. Existing FD methods generally treat these models equally by merely computing the average of their output soft predictions for some given input distillation sample, which does not take the diversity across all local models into account, thus leading to degraded performance of the aggregated model, especially when some local models learn little knowledge about the sample. In this paper, we propose a new perspective that treats the local data in each client as a specific domain and design a novel domain knowledge aware federated distillation method, dubbed DaFKD, that can discern the importance of each model to the distillation sample, and thus is able to optimize the ensemble of soft predictions from diverse models. Specifically, we employ a domain discriminator for each client, which is trained to identify the correlation factor between the sample and the corresponding domain. Then, to facilitate the training of the domain discriminator while saving communication costs, we propose sharing its partial parameters with the classification model. Extensive experiments on various datasets and settings show that the proposed method can improve the model accuracy by up to 6.02% compared to state-of-the-art baselines.

Federated Learning (FL) allows massive clients to collaboratively train a global model without revealing their private data. Because of the participants’ not independently and identically distributed (non-IID) statistical characteristics, it will cause divergence among the client’s Deep Neural Network model weights and require more communication rounds before training can be converged. Moreover, models trained from non-IID data may also extract biased features and the rationale behind the model is still not fully analyzed and exploited. In this paper, we propose eXplainable-Fed (XFed) which is a novel client selection mechanism that takes both accuracy and explainability into account. Specifically, XFed selects participants in each round based on a small test set’s accuracy via cross-entropy loss and interpretability via XAI-accuracy. XAI-accuracy is calculated by Intersection over Union Ratio between the heat map and the truth mask to evaluate the overall rationale of accuracy. The results of our experiments show that our method has comparable accuracy to state-of-the-art methods specially designed for accuracy while increasing explainability by 14%-35% in terms of rationality.

Federated Learning enables collaboratively model training among a number of distributed devices with the coordination of a centralized server, where each device alternatively performs local gradient computation and communication to the server. FL suffers from significant performance degradation due to the excessive communication delay between the server and devices, especially when the network bandwidth of these devices is limited, which is common in edge environments. Existing methods overlap the gradient computation and communication to hide the communication latency to accelerate the FL training. However, the overlapping can also lead to an inevitable gap between the local model in each device and the global model in the server that seriously restricts the convergence rate of learning process. To address this problem, we propose a new overlapping method for FL, AOCC-FL, which aligns the local model with the global model via calibrated compensation such that the communication delay can be hidden without deteriorating the convergence performance. Theoretically, we prove that AOCC-FL admits the same convergence rate as the non-overlapping method. On both simulated and testbed experiments, we show that AOCC-FL achieves a comparable convergence rate relative to the non-overlapping method while outperforming the state-of-the-art overlapping methods.

A self-explaining rationalization model is generally constructed by a cooperative game where a generator selects the most human-intelligible pieces from the input text as rationales, followed by a predictor that makes predictions based on the selected rationales. However, such a cooperative game may incur the degeneration problem where the predictor overfits to the uninformative pieces generated by a not yet well-trained generator and in turn, leads the generator to converge to a sub-optimal model that tends to select senseless pieces. In this paper, we theoretically bridge degeneration with the predictor's Lipschitz continuity. Then, we empirically propose a simple but effective method named DR, which can naturally and flexibly restrain the Lipschitz constant of the predictor, to address the problem of degeneration. The main idea of DR is to decouple the generator and predictor to allocate them with asymmetric learning rates. A series of experiments conducted on two widely used benchmarks have verified the effectiveness of the proposed method. Codes: https://github.com/jugechengzi/Rationalization-DR.

Rationalization aims to strengthen the interpretability of NLP models by extracting a subset of human-intelligible pieces of their inputting texts. Conventional works generally employ the maximum mutual information (MMI) criterion to find the rationale that is most indicative of the target label. However, this criterion can be influenced by spurious features that correlate with the causal rationale or the target label. Instead of attempting to rectify the issues of the MMI criterion, we propose a novel criterion to uncover the causal rationale, termed the Minimum Conditional Dependence (MCD) criterion, which is grounded on our finding that the non-causal features and the target label are \emph{d-separated} by the causal rationale. By minimizing the dependence between the non-selected parts of the input and the target label conditioned on the selected rationale candidate, all the causes of the label are compelled to be selected. In this study, we employ a simple and practical measure for dependence, specifically the KL-divergence, to validate our proposed MCD criterion. Empirically, we demonstrate that MCD improves the F1 score by up to 13.7% compared to previous state-of-the-art MMI-based methods. Our code is in an anonymous repository: https://anonymous.4open.science/r/MCD-CE88.

In the quest to advance human-centric natural language generation (NLG) systems, ensuring alignment between NLG models and human preferences is crucial. For this alignment, current popular methods leverage a reinforcement learning (RL) approach with a reward model trained on feedback from humans. However, inherent disagreements due to the subjective nature of human preferences pose a significant challenge for training the reward model, resulting in a deterioration of the NLG performance. To tackle this issue, previous approaches typically rely on majority voting or averaging to consolidate multiple inconsistent preferences into a merged one. Although straightforward to understand and execute, such methods suffer from an inability to capture the nuanced degrees of disaggregation among humans and may only represent a specialized subset of individuals, thereby lacking the ability to quantitatively disclose the universality of human preferences. To address this challenge, this paper proposes a novel approach, which employs a Bayesian framework to account for the distribution of disagreements among human preferences as training a preference model, and names it as $\textbf{d-PM}$. Besides, considering the RL strategy's inefficient and complex training process over the training efficiency, we further propose utilizing the contrastive learning strategy to train the NLG model with the preference scores derived from the d-PM model. Extensive experiments on two human-centric NLG tasks, i.e., emotional support conversation and integrity ``Rule-of-Thumb'' generation, show that our method consistently exceeds previous SOTA models in both automatic and human evaluations.

Temporal sentence grounding (TSG) is an important yet challenging task in video-based information retrieval. Given an untrimmed video input, it requires the machine to predict the interested video segment semantically related to a given sentence query. Most existing TSG methods train well-designed deep networks to align the semantic between video-query pairs for activity grounding with a large amount of data. However, we argue that these works easily capture the selection biases of video-query pairs in a dataset rather than showing the robust reasoning abilities to handle the rarely appeared pairs (i.e., few-shot contents). To alleviate such limitation of the off-balance data distribution during the network training, in this paper, we propose a novel memory-augmented network called Memory-Guided Semantic Learning Network (MGSL-Net) to handle the few-shot TSG task for enhancing the model generalization ability. Specifically, given the matched video-query input, we first employ a graph attentive cross-modal interaction module to align their semantics in a cycle-consistent manner. Then, we develop the memory modules in both video and query domains to record the cross-modal shared semantic features in the domain-specific persistent memory. At last, a heterogeneous attention module is utilized to integrate the memory-enhanced multi-modal features in both video and query domains with further feature calibration. During training, the memory modules are dynamically associated with both common and rare cases to memorize all appeared contents, alleviating the issue of forgetting the few-shot contents. Therefore, in testing, the rare cases can be enhanced by retrieving the stored memories, improving the generalization ability of the model. Experimental results on three benchmarks (ActivityNet Caption, Charades-STA and TACoS) show the superiority of our method on both effectiveness and efficiency.

In the setting of federated optimization, where a global model is aggregated periodically, step asynchronism occurs when participants conduct model training by efficiently utilizing their computational resources. It is well acknowledged that step asynchronism leads to objective inconsistency under non-i.i.d. data, which degrades the model’s accuracy. To address this issue, we propose a new algorithmFedaGrac, which calibrates the local direction to a predictive global orientation. Taking advantage of the estimated orientation, we guarantee that the aggregated model does not excessively deviate from the global optimum while fully utilizing the local updates of faster nodes. We theoretically prove thatFedaGracholds an improved order of convergence rate than the state-of-the-art approaches and eliminates the negative effect of step asynchronism. Empirical results show that our algorithm accelerates the training and enhances the final accuracy.

Traditional one-bit compressed stochastic gradient descent can not be directly employed in multi-hop all-reduce, a widely adopted distributed training paradigm in network-intensive high-performance computing systems such as public clouds. According to our theoretical findings, due to the cascading compression, the training process has considerable deterioration on the convergence performance. To overcome this limitation, we implement a sign-bit compression-based learning synchronization framework, Marsit. It prevents cascading compression via an elaborate bit-wise operation for unbiased sign aggregation and its specific global compensation mechanism for mitigating compression deviation. The proposed framework retains the same theoretical convergence rate as non-compression mechanisms. Experimental results demonstrate that Marsit reduces up to 35% training time while preserving the same accuracy as training without compression.

Rationalization aims to strengthen the interpretability of NLP models by extracting a subset of human-intelligible pieces of their inputting texts. Conventional works generally employ a two-phase model in which a generator selects the most important pieces, followed by a predictor that makes predictions based on the selected pieces. However, such a two-phase model may incur the degeneration problem where the predictor overfits to the noise generated by a not yet well-trained generator and in turn, leads the generator to converge to a suboptimal model that tends to select senseless pieces. To tackle this challenge, we propose Folded Rationalization (FR) that folds the two phases of the rationale model into one from the perspective of text semantic extraction. The key idea of FR is to employ a unified encoder between the generator and predictor, based on which FR can facilitate a better predictor by access to valuable information blocked by the generator in the traditional two-phase model and thus bring a better generator. Empirically, we show that FR improves the F1 score by up to 10.3% as compared to state-of-the-art methods.

Federated Learning has attracted widespread attention in recent years because it allows massive edge nodes to collaboratively train machine learning models without sharing their private data sets. However, these edge nodes are usually heterogeneous in computational capability and statistically different in data distribution, i.e., non-independent and identically distributed (IID), leading to significant performance degradation. Although status quo asynchronous training methods can solve the heterogeneity issue, they cannot prevent the non-IID problem from reducing the convergence rate. In this article, we propose a novel paradigm that schedules the gradient with partially averaged gradients and applies the global momentum (GSGM) for asynchronous training over non-IID data sets in an edge environment. Our key idea is to apply global momentum and partial average on the biased gradients calculated on edge nodes after scheduling, to make the training process stable. Empirical results demonstrate that GSGM can well adapt to different degrees of non-IID data and bring 20% performance gains in terms of training stability for popular optimization algorithms with enhanced accuracy over Fashion-Mnist and CIFAR-10 data sets.

The ever-growing artificial intelligence (AI) applications have greatly reshaped our world in many areas, e.g., smart home, computer vision, natural language processing, etc. Behind these applications are usually machine learning (ML) models with extremely large size, which require huge data sets for accurate training to mine the value contained in the big data. Large ML models, however, can consume tremendous computing resources to achieve decent performance and thus, it is difficult to train them in resource-constrained Internet of Things (IoT) environments, which would prevent further development and application of AI techniques in the future. To deal with such challenges, there are many efforts on accelerating the training process for large ML models in IoT. In this article, we provide a comprehensive review on the recent advances toward reducing the computing cost during the training stage while maintaining comparable model accuracy. Specifically, the optimization algorithms that aim to improve the convergence rate are emphasized over various distributed learning architectures that exploit ubiquitous computing resources. Then, the article elaborates the computation hardware acceleration and communication optimization for collaborative training among multiple learning entities. Finally, the remaining challenges, future opportunities, and possible directions are discussed.

Distributed gradient descent has been widely adopted in the machine learning field because considerable computing resources are available when facing the huge volume of data. Specifically, the gradient over the whole data is cooperatively computed by multiple workers. However, its performance can be severely affected by slow workers, namely stragglers. Recently, coding-based approaches have been introduced to mitigate the straggler problem, but they could hardly deal with the heterogeneity among workers. Besides, they always discard the results of stragglers causing huge resource waste. In this article, we first investigate how to tolerate stragglers by discarding their results and then seek to leverage the stragglers. For tolerating stragglers, we propose a heterogeneity-aware coding scheme that encodes gradients adaptive to the computing capability of workers. Theoretically, this scheme is optimal for stragglers tolerance. Relying on the scheme, we further propose an algorithm called DHeter-aware to exploit the gradients of stragglers which we called delayed gradients. Moreover, theoretical results characterized for DHeter-aware exhibits the same convergence rate as the gradient descent without delayed gradients. Experiments on various tasks and clusters demonstrate that our coding scheme outperforms all the state-of-the-art methods and the DHeter-aware further accelerates the coding scheme by achieving 25 percent time savings.

Federated Learning (FL) is a promising machine learning paradigm to cooperatively train a global model with highly distributed data located on mobile devices. Aiming to optimize the communication efficiency for gradient aggregation and model synchronization among large-scale devices, we propose a relay-assisted FL framework. By breaking the traditional transmission-order constraint and exploiting the broadcast characteristic of relay nodes, we design a novel synchronization scheme named Partial Synchronization Parallel (PSP), in which models and gradients are transmitted simultaneously and aggregated at relay nodes, resulting in traffic reduction. We prove that PSP has the same convergence rate as the sequential synchronization approaches via rigorous analysis. To further accelerate the training process, we integrate PSP with any unbiased and error-bounded compression technologies and prove that the convergence properties of the resulting scheme still hold. Extensive experiments are conducted in a distributed cluster environment with real-world datasets and the results demonstrate that our proposed approach reduces the training time up to 37 percent compared to state-of-the-art methods.

Communication has been considered as a major bottleneck in large-scale decentralized training systems since participating nodes iteratively exchange large amounts of intermediate data with their neighbors. Although compression techniques like sparsification can significantly reduce the communication overhead in each iteration, errors caused by compression will be accumulated, resulting in a severely degraded convergence rate. Recently, the error compensation method for sparsification has been proposed in centralized training to tolerate the accumulated compression errors. However, the analog technique and the corresponding theory about its convergence in decentralized training are still unknown. To fill in the gap, we design a method named ECSD-SGD that significantly accelerates decentralized training via error-compensated sparsification. The novelty lies in that we identify the component of the exchanging information in each iteration (i.e., the sparsified model update) and make targeted error compensation over the component. Our thorough theoretical analysis shows that ECSD-SGD supports arbitrary sparsification ratio and achieves the same convergence rate as the non-sparsified decentralized training methods. We also conduct extensive experiments on multiple deep learning models to validate our theoretical findings. Results show that ECSD-SGD outperforms all the start-of-the-art sparsified methods in terms of both the convergence speed and the final generalization accuracy.

Most of the existing FL systems focus on a data-parallel architecture where training data are partitioned by samples among several parties. In some real-life applications, however, partitioning by features is also of practical relevance and the number of features is usually unbalanced among parties. The corresponding learning framework is referred to as Vertical Federated Learning (VFL). Though some pioneering work focused on VFL, the convergence properties of VFL on unbalanced features, especially when parties conduct different numbers of local updates concerning heterogeneous computational capabilities are still unknown. In this article, we propose a new learning framework to improve the training efficiency of VFL on unbalanced features. Given the number of features and the computational capability owned by each party, our thorough theoretical analysis exhibits that the number of local updates conducted by each party has a great effect on the convergence rate and the computational complexity, both of which jointly determine the overall training efficiency in an interrelated and sophisticated way. Based on our theoretical findings, we formulate an optimization problem and derive the optimal solution by selecting an adaptive number of local training rounds for each party. Extensive experiments on various datasets and models demonstrate that our approach significantly improves the training efficiency of VFL.

In recent years, personalized federated learning (pFL) has attracted increasing attention for its potential in dealing with statistical heterogeneity among clients. However, the state-of-the-art pFL methods rely on model parameters aggregation at the server side, which require all models to have the same structure and size, and thus limits the application for more heterogeneous scenarios. To deal with such model constraints, we exploit the potentials of heterogeneous model settings and propose a novel training framework to employ personalized models for different clients. Specifically, we formulate the aggregation procedure in original pFL into a personalized group knowledge transfer training algorithm, namely, KT-pFL, which enables each client to maintain a personalized soft prediction at the server side to guide the others' local training. KT-pFL updates the personalized soft prediction of each client by a linear combination of all local soft predictions using a knowledge coefficient matrix, which can adaptively reinforce the collaboration among clients who own similar data distribution. Furthermore, to quantify the contributions of each client to others' personalized training, the knowledge coefficient matrix is parameterized so that it can be trained simultaneously with the models. The knowledge coefficient matrix and the model parameters are alternatively updated in each round following the gradient descent way. Extensive experiments on various datasets (EMNIST, Fashion_MNIST, CIFAR-10) are conducted under different settings (heterogeneous models and data distributions). It is demonstrated that the proposed framework is the first federated learning paradigm that realizes personalized model training via parameterized group knowledge transfer while achieving significant performance gain comparing with state-of-the-art algorithms.

Spelling Error Correction (SEC) that detects and corrects spelling errors in a text has a wide range of applications in human language understanding. Earlier solutions, including statistic-based methods, one-stage, and two-stage machine learning-based methods, cannot build deeply bidirectional models and significantly confine the learning ability. With the recently emerging masked language models, transformer-based networks have achieved remarkable success in SEC. However, current transformer-based Chinese SEC algorithms are all end-to-end methods, which suffer from high false alarm rates because they correct each character of the sentence regardless of its correctness. This issue becomes even more severe when there exist only a small fraction of incorrect characters in the whole sentence. To solve this problem, we propose a cloze-style detector-corrector framework (DCSpell) that firstly detects whether a character is erroneous before correcting it. Specifically, DCSpell employs the discriminator of ELECTRA as the Detector to detect the positions of incorrect characters. The Detector is trained by a sample-efficient replaced token detection pre-training task, and thus allows domain adaption with a small amount of data. After that, a transformer-based Corrector is used to find the correct character for each detected position. It employs sentence pairs as the input, which potentially incorporates the knowledge of phonological and visual similarity. A confusion-set-based post-processing is used to further improve the performance. Experiments show that DCSpell achieves 15.7% improvement on the SIGHAN dataset and 6.6% improvement on a dataset transcribed from a real-world acoustic speech corpus compared to the state-of-the-art methods in terms of the F1 score.

When running in Parameter Server (PS), the Distributed Stochastic Gradient Descent (D-SGD) incurs significant communication delays and huge communication overhead due to the model synchronization. Moreover, considering the heterogeneity of computational capability among workers, traditional synchronization modes incur under-utilization of computational resources because fast workers have to wait for slow ones finishing the computation. Although our previous work OSP can effectively solve these problems by overlapping the computation and communication procedures and allowing adaptive multiple local updates in distributed training, it causes the staleness problem brought by the overlap, yielding a performance degradation. In this paper, we propose a new method named LOSP by introducing local compensation to our previous synchronization mechanism, which mitigates adverse effects caused by the overlapping synchronization. We theoretically prove that LOSP (1) preserves the same convergence rate as the sequential SGD for non-convex problems, and (2) exhibits good scalability due to the linear speedup property with respect to both the number of workers and the average number of local updates. Evaluations show that LOSP significantly improves performance over the state-of-the-art ones in terms of both convergence accuracy and communication cost.

Gradient descent algorithms are widely used in machine learning. In order to deal with huge volume of data, we consider the implementation of gradient descent algorithms in a distributed computing setting where multiple workers compute the gradient over some partial data and the master node aggregates their results to obtain the gradient over the whole data. However, its performance can be severely affected by straggler workers. Recently, some coding-based approaches are introduced to mitigate the straggler problem, but they are efficient only when the workers are homogeneous, i.e., having the same computation capabilities. In this paper, we consider that the workers are heterogenous which are common in modern distributed systems. We propose a novel heterogeneity-aware gradient coding scheme which can not only tolerate a predetermined number of stragglers but also fully utilize the computation capabilities of heterogenous workers. We show that this scheme is optimal when the computation capabilities of workers are estimated accurately. A variant of this scheme is further proposed to improve the performance when the estimations of the computation capabilities are not so accurate. We conduct our schemes for gradient descent based image classification on QingCloud clusters. Evaluation results show that our schemes can reduce the whole computation time by up to 3× compared with a state-of-the-art coding scheme.

When running in Parameter Server (PS), the Distributed Stochastic Gradient Descent (SGD) incurs significant communication delays because after pushing their updates, computing nodes (workers) have to wait for the global model to be communicated back from the master in every iteration. In this paper, we devise a new synchronization parallel mechanism named overlap synchronization parallel (OSP), in which the waiting time is removed by conducting computation and communication in an overlapped manner. We theoretically prove that our mechanism could achieve the same convergence rate compared to the sequential SGD for non-convex problems. Evaluations show that our mechanism significantly improves performance over the state-of-the-art ones, e.g., by 4× for both AlexNet and ResNet18 in terms of convergence speed.