Predicting the tensor properties of crystalline materials is a fundamental task in materials science. Unlike single-value property prediction, which is inherently invariant, tensor property prediction requires maintaining O(3) group tensor equivariance. Such equivariance constraint often requires specialized architecture designs to achieve effective predictions, inevitably introducing tremendous computational costs. Canonicalization, a classical technique for geometry, has recently been explored for efficient learning with symmetry. In this work, we revisit the problem of crystal tensor property prediction through the lens of canonicalization. Specifically, we demonstrate how polar decomposition, a simple yet efficient algebraic method, can serve as a form of canonicalization and be leveraged to ensure equivariant tensor property prediction. Building upon this insight, we propose a general O(3)-equivariant framework for efficient crystal tensor property prediction, referred to as GoeCTP. By utilizing canonicalization, GoeCTP achieves high efficiency without requiring the explicit incorporation of equivariance constraints into the network architecture. Experimental results indicate that GoeCTP achieves the best prediction performance and runs at most 13 times faster compared to existing state-of-the-art methods in benchmarking datasets, underscoring its effectiveness and efficiency.

Text-to-motion generation, which synthesizes 3D human motions from text inputs, holds immense potential for applications in gaming, film, and robotics. Recently, diffusion-based methods have been shown to generate more diversity and realistic motion. However, there exists a misalignment between text and motion distributions in diffusion models, which leads to semantically inconsistent or low-quality motions. To address this limitation, we propose Reward-guided sampling Alignment (ReAlign), comprising a step-aware reward model to assess alignment quality during the denoising sampling and a reward-guided strategy that directs the diffusion process toward an optimally aligned distribution. This reward model integrates step-aware tokens and combines a text-aligned module for semantic consistency and a motion-aligned module for realism, refining noisy motions at each timestep to balance probability density and alignment. Extensive experiments of both motion generation and retrieval tasks demonstrate that our approach significantly improves text-motion alignment and motion quality compared to existing state-of-the-art methods.

Recent approaches attempt to adapt powerful interactive segmentation models, such as SAM, to interactive matting and fine-tune the models based on synthetic matting datasets. However, models trained on synthetic data fail to generalize to complex and occlusion scenes. We address this challenge by proposing a new matting dataset based on the COCO dataset, namely COCO-Matting. It selects real-world complex images from COCO and converts semantic segmentation masks to matting labels. The built COCO-Matting comprises an extensive collection of 36,980 human instance-level alpha mattes in complex natural scenarios. Furthermore, existing SAM-based matting methods extract intermediate features and masks from a frozen SAM and only train a lightweight matting decoder by end-to-end matting losses, which do not fully exploit the potential of the pre-trained SAM. Thus, we propose SEMat which revamps the network architecture and training objectives. For network architecture, the proposed feature-aligned transformer learns to extract fine-grained edge and transparency features. The proposed matte-aligned decoder aims to segment matting-specific objects and convert coarse masks into high-precision mattes. For training objectives, the proposed regularization and trimap loss aim to retain the prior from the pre-trained model and push the matting logits extracted from the mask decoder to contain trimap-based semantic information. Extensive experiments across seven diverse datasets demonstrate the superior performance of our method, proving its efficacy in interactive natural image matting. Code is available at https://github.com/XiaRho/SEMat.

Embodied agents based on large language models (LLMs) face significant challenges in collaborative tasks, requiring effective communication and reasonable division of labor to ensure efficient and correct task completion. Previous approaches with simple communication patterns carry erroneous or incoherent agent actions, which can lead to additional risks. To address these problems, we propose Cooperative Tree Search (CoTS), a framework designed to significantly improve collaborative planning and task execution efficiency among embodied agents. CoTS guides multi-agents to discuss long-term strategic plans within a modified Monte Carlo tree, searching along LLM-driven reward functions to provide a more thoughtful and promising approach to cooperation. Another key feature of our method is the introduction of a plan evaluation module, which not only prevents agent action confusion caused by frequent plan updates but also ensures plan updates when the current plan becomes unsuitable. Experimental results show that the proposed method performs excellently in planning, communication, and collaboration on embodied environments (CWAH and TDW-MAT), efficiently completing long-term, complex tasks and significantly outperforming existing methods.

Preconditioned stochastic optimization algorithms, exemplified by Shampoo, outperform first-order optimizers by offering theoretical convergence benefits and practical gains in large-scale neural network training. However, they incur substantial memory overhead due to the storage demands of non-diagonal preconditioning matrices. To address this, we introduce 4-bit quantization for Shampoo's preconditioners. We introduce two key methods: First, we apply Cholesky decomposition followed by quantization of the Cholesky factors, reducing memory usage by leveraging their lower triangular structure while better preserving spectral properties to minimize information loss. To our knowledge, this is the first quantization approach applied to Cholesky factors of preconditioners. Second, we incorporate error feedback in the quantization process, efficiently storing Cholesky factor and error state in the lower and upper triangular parts of the same matrix. Through extensive experiments, we demonstrate that combining Cholesky quantization with error feedback enhances memory efficiency and algorithm performance in large-scale deep-learning tasks. Theoretically, we also provide convergence proofs for quantized Shampoo under both smooth and non-smooth stochastic optimization settings.

Generalized Few-Shot Semantic Segmentation (GFSS) aims to extend a segmentation model to novel classes with only a few annotated examples while maintaining performance on base classes. Recently, pretrained vision-language models (VLMs) such as CLIP have been leveraged in GFSS to improve generalization on novel classes through multi-modal prototypes learning. However, existing prototype-based methods are inherently deterministic, limiting the adaptability of learned prototypes to diverse samples, particularly for novel classes with scarce annotations. To address this, we propose FewCLIP, a probabilistic prototype calibration framework over multi-modal prototypes from the pretrained CLIP, thus providing more adaptive prototype learning for GFSS. Specifically, FewCLIP first introduces a prototype calibration mechanism, which refines frozen textual prototypes with learnable visual calibration prototypes, leading to a more discriminative and adaptive representation. Furthermore, unlike deterministic prototype learning techniques, FewCLIP introduces distribution regularization over these calibration prototypes. This probabilistic formulation ensures structured and uncertainty-aware prototype learning, effectively mitigating overfitting to limited novel class data while enhancing generalization. Extensive experimental results on PASCAL-5$^i$ and COCO-20$^i$ datasets demonstrate that our proposed FewCLIP significantly outperforms state-of-the-art approaches across both GFSS and class-incremental setting. The code is available at https://github.com/jliu4ai/FewCLIP.

Semi-supervised learning (SSL), exemplified by FixMatch (Sohn et al., 2020), has shown significant generalization advantages over supervised learning (SL), particularly in the context of deep neural networks (DNNs). However, it is still unclear, from a theoretical standpoint, why FixMatch-like SSL algorithms generalize better than SL on DNNs. In this work, we present the first theoretical justification for the enhanced test accuracy observed in FixMatch-like SSL applied to DNNs by taking convolutional neural networks (CNNs) on classification tasks as an example. Our theoretical analysis reveals that the semantic feature learning processes in FixMatch and SL are rather different. In particular, FixMatch learns all the discriminative features of each semantic class, while SL only randomly captures a subset of features due to the well-known lottery ticket hypothesis. Furthermore, we show that our analysis framework can be applied to other FixMatch-like SSL methods, e.g., FlexMatch, FreeMatch, Dash, and SoftMatch. Inspired by our theoretical analysis, we develop an improved variant of FixMatch, termed Semantic-Aware FixMatch (SA-FixMatch). Experimental results corroborate our theoretical findings and the enhanced generalization capability of SA-FixMatch.

In this work, we address the cooperation problem among large language model (LLM) based embodied agents, where agents must cooperate to achieve a common goal. Previous methods often execute actions extemporaneously and incoherently, without long-term strategic and cooperative planning, leading to redundant steps, failures, and even serious repercussions in complex tasks like search-and-rescue missions where discussion and cooperative plan are crucial. To solve this issue, we propose Cooperative Plan Optimization (CaPo) to enhance the cooperation efficiency of LLM-based embodied agents. Inspired by human cooperation schemes, CaPo improves cooperation efficiency with two phases: 1) meta plan generation, and 2) progress-adaptive meta plan and execution. In the first phase, all agents analyze the task, discuss, and cooperatively create a meta-plan that decomposes the task into subtasks with detailed steps, ensuring a long-term strategic and coherent plan for efficient coordination. In the second phase, agents execute tasks according to the meta-plan and dynamically adjust it based on their latest progress (e.g., discovering a target object) through multi-turn discussions. This progress-based adaptation eliminates redundant actions, improving the overall cooperation efficiency of agents. Experimental results on the ThreeDworld Multi-Agent Transport and Communicative Watch-And-Help tasks demonstrate CaPo's much higher task completion rate and efficiency compared with state-of-the-arts. The code is released at https://github.com/jliu4ai/CaPo.

Interactive 3D segmentation has emerged as a promising solution for generating accurate object masks in complex 3D scenes by incorporating user-provided clicks. However, two critical challenges remain underexplored: (1) effectively generalizing from sparse user clicks to produce accurate segmentations and (2) quantifying predictive uncertainty to help users identify unreliable regions. In this work, we propose \emph{NPISeg3D}, a novel probabilistic framework that builds upon Neural Processes (NPs) to address these challenges. Specifically, NPISeg3D introduces a hierarchical latent variable structure with scene-specific and object-specific latent variables to enhance few-shot generalization by capturing both global context and object-specific characteristics. Additionally, we design a probabilistic prototype modulator that adaptively modulates click prototypes with object-specific latent variables, improving the model’s ability to capture object-aware context and quantify predictive uncertainty. Experiments on four 3D point cloud datasets demonstrate that NPISeg3D achieves superior segmentation performance with fewer clicks while providing reliable uncertainty estimations.

Aligning Large Language Model (LLM) responses with human preferences is vital for building safe and controllable AI systems. While preference optimization methods based on Plackett-Luce (PL) and Bradley-Terry (BT) models have shown promise, they face challenges such as poor handling of harmful content, inefficient use of dispreferred responses, and, specifically for PL, high computational costs. To address these issues, we propose Hard Preference Sampling (HPS), a novel framework for robust and efficient human preference alignment. HPS introduces a training loss that prioritizes the most preferred response while rejecting all dispreferred and harmful ones. It emphasizes “hard” dispreferred responses — those closely resembling preferred ones — to enhance the model’s rejection capabilities. By leveraging a single-sample Monte Carlo sampling strategy, HPS reduces computational overhead while maintaining alignment quality. Theoretically, HPS improves sample efficiency over existing PL methods and maximizes the reward margin between preferred and dispreferred responses, ensuring clearer distinctions. Experiments on HH-RLHF and PKU-Safety datasets validate HPS’s effectiveness, achieving comparable BLEU and reward scores while greatly improving reward margins and thus reducing harmful content generation.

Speculative decoding accelerates inference in large language models (LLMs) by generating multiple draft tokens simultaneously. However, existing methods often struggle with token misalignment between the training and decoding phases, limiting their performance. To address this, we propose GRIFFIN, a novel framework that incorporates a token-alignable training strategy and a token-alignable draft model to mitigate misalignment. The training strategy employs a loss masking mechanism to exclude highly misaligned tokens during training, preventing them from negatively impacting the draft model's optimization. The token-alignable draft model introduces input tokens to correct inconsistencies in generated features. Experiments on LLaMA, Vicuna, Qwen and Mixtral models demonstrate that GRIFFIN achieves an average acceptance length improvement of over 8\% and a speedup ratio exceeding 7\%, outperforming current speculative decoding state-of-the-art methods. Our code and GRIFFIN's draft models will be released publicly in https://github.com/hsj576/GRIFFIN.

Text-to-motion generation is essential for advancing the creative industry but often presents challenges in producing consistent, realistic motions. To address this, we focus on fine-tuning text-to-motion models to consistently favor high-quality, human-preferred motions—a critical yet largely unexplored problem. In this work, we theoretically investigate the DPO under both online and offline settings, and reveal their respective limitation: overfitting in offline DPO, and biased sampling in online DPO. Building on our theoretical insights, we introduce Semi-online Preference Optimization (SoPo), a DPO-based method for training text-to-motion models using ``semi-online” data pair, consisting of unpreferred motion from online distribution and preferred motion in offline datasets. This method leverages both online and offline DPO, allowing each to compensate for the other’s limitations. Extensive experiments demonstrate that SoPo outperforms other preference alignment methods, with an MM-Dist of 3.25\% (vs e.g. 0.76\% of MoDiPO) on the MLD model, 2.91\% (vs e.g. 0.66\% of MoDiPO) on MDM model, respectively. Additionally, the MLD model fine-tuned by our SoPo surpasses the SoTA model in terms of R-precision and MM Dist. Visualization results also show the efficacy of our SoPo in preference alignment. Project page: https://xiaofeng-tan.github.io/projects/SoPo/.

Recently, numerous benchmarks have been developed to evaluate the logical reasoning abilities of large language models (LLMs). However, assessing the equally important creative capabilities of LLMs is challenging due to the subjective, diverse, and data-scarce nature of creativity, especially in multimodal scenarios. In this paper, we consider the comprehensive pipeline for evaluating the creativity of multimodal LLMs, with a focus on suitable evaluation platforms and methodologies. First, we find the Oogiri game-a creativity-driven task requiring humor, associative thinking, and the ability to produce unexpected responses to text, images, or both. This game aligns well with the input-output structure of modern multimodal LLMs and benefits from a rich repository of high-quality, human-annotated creative responses, making it an ideal platform for studying LLM creativity. Next, beyond using the Oogiri game for standard evaluations like ranking and selection, we propose LoTbench, an interactive, causality-aware evaluation framework, to further address some intrinsic risks in standard evaluations, such as information leakage and limited interpretability. The proposed LoTbench not only quantifies LLM creativity more effectively but also visualizes the underlying creative thought processes. Our results show that while most LLMs exhibit constrained creativity, the performance gap between LLMs and humans is not insurmountable. Furthermore, we observe a strong correlation between results from the multimodal cognition benchmark MMMU and LoTbench, but only a weak connection with traditional creativity metrics. This suggests that LoTbench better aligns with human cognitive theories, highlighting cognition as a critical foundation in the early stages of creativity and enabling the bridging of diverse concepts. Project Page.

To efficiently train large-scale models, low-bit gradient communication compresses full-precision gradients on local GPU nodes into low-precision ones for higher gradient synchronization efficiency among GPU nodes. However, it often degrades training quality due to compression information loss. To address this, we propose the Low-bit Communication Adaptor (LoCo), which compensates gradients on local GPU nodes before compression, ensuring efficient synchronization without compromising training quality. Specifically, LoCo designs a moving average of historical compensation errors to stably estimate concurrent compression error and then adopts it to compensate for the concurrent gradient compression, yielding a less lossless compression. This mechanism allows it to be compatible with general optimizers like Adam and sharding strategies like FSDP. Theoretical analysis shows that integrating LoCo into full-precision optimizers like Adam and SGD does not impair their convergence speed on non-convex problems. Experimental results show that across large-scale model training frameworks like Megatron-LM and PyTorch's FSDP, LoCo significantly improves communication efficiency, e.g., improving Adam's training speed by 14% to 40% without performance degradation on large language models like LLAMAs and MoEs.

Sharpness-Aware Minimization (SAM) has been instrumental in improving deep neural network training by minimizing both training loss and loss sharpness. Despite the practical success, the mechanisms behind SAM's generalization enhancements remain elusive, limiting its progress in deep learning optimization. In this work, we investigate SAM's core components for generalization improvement and introduce “Friendly-SAM” (F-SAM) to further enhance SAM's generalization. Our investigation reveals the key role of batch-specific stochastic gradient noise within the adversarial perturbation, i.e., the current minibatcli gradient, which significantly influences SAM's generalization performance. By decomposing the adversarial perturbation in SAM into full gradient and stochastic gradient noise components, we discover that relying solely on the full gradient component degrades generalization while excluding it leads to improved performance. The possible reason lies in the full gradient component's increase in sharpness loss for the entire dataset, creating inconsistencies with the subsequent sharpness minimization step solely on the current minibatcli data. Inspired by these insights, F-SAM aims to mitigate the negative effects of the full gradient component. It removes the full gradient estimated by an exponentially moving average (EMA) of historical stochastic gradients, and then leverages stochastic gradient noise for improved generalization. Moreover, we provide theoretical validation for the EMA approximation and prove the convergence of F-SAM on non-convex problems. Extensive experiments demonstrate the superior generalization performance and robustness of F-SAM over vanilla SAM. Code is available at https://github.com/nblt/F-SAM.

Score distillation sampling (SDS) and its variants have greatly boosted the development of text-to-3D generation, but are vulnerable to geometry collapse and poor textures yet. To solve this issue, we first deeply analyze the SDS and find that its distillation sampling process indeed corresponds to the trajectory sampling of a stochastic differential equation (SDE): SDS samples along an SDE trajectory to yield a less noisy sample which then serves as a guidance to optimize a 3D model. However, the randomness in SDE sampling often leads to a diverse and unpredictable sample which is not always less noisy, and thus is not a consistently correct guidance, explaining the vulnerability of SDS. Since for any SDE, there always exists an ordinary differential equation (ODE) whose trajectory sampling can deterministically and consistently converge to the desired target point as the SDE, we propose a novel and effective “Consistent3D” method that explores the ODE deterministic sampling prior for text-to-3D generation. Specifically, at each training iteration, given a rendered image by a 3D model, we first estimate its desired 3D score function by a pre-trained 2D diffusion model, and build an ODE for trajectory sampling. Next, we design a consistency distillation sampling loss which samples along the ODE trajectory to generate two adjacent samples and uses the less noisy sample to guide another more noisy one for distilling the deterministic prior into the 3D model. Experimental results show the efficacy of our Consistent3D in generating high-fidelity and diverse 3D objects and large-scale scenes, as shown in Fig. 1. The codes are available at https://github.com/sail-sg/Consistent3D.

Score distillation sampling (SDS) has been widely adopted to overcome the absence of unseen views in reconstructing 3D objects from a single image. It leverages pretrained 2D diffusion models as teacher to guide the reconstruction of student 3D models. Despite their remarkable success, SDS-based methods often encounter geometric artifacts and texture saturation. We find out the crux is the overlooked indiscriminate treatment of diffusion time-steps during optimization: it unreasonably treats the student-teacher knowledge distillation to be equal at all time-steps and thus entangles coarse-grained and fine-grained modeling. Therefore, we propose the Diffusion Time-step Curriculum one-image-to-3D pipeline (DTC123), which involves both the teacher and student models collaborating with the time-step curriculum in a coarse-to-fine manner. Extensive experiments on NeRF4, RealFusion15, GSO and Level50 benchmark demonstrate that DTC123 can produce multiview consistent, high-quality, and diverse 3D assets. Codes and more generation demos will be released in https://github.com/yxymessi/DTC123.

Inspired by the long-range modeling ability of ViTs, large-kernel convolutions are widely studied and adopted recently to enlarge the receptive field and improve model performance, like the remarkable work ConvNeXt which employs 7 × 7 depthwise convolution. Although such depth-wise operator only consumes a few FLOPs, it largely harms the model efficiency on powerful computing devices due to the high memory access costs. For example, ConvNeXt-T has similar FLOPs with ResNet-50 but only achieves ~60% throughputs when trained on A100 GPUs with full precision. Although reducing the kernel size of ConvNeXt can improve speed, it results in significant performance degradation, which poses a challenging problem: How to speed up large-kernel-based CNN models while preserving their performance. To tackle this issue, inspired by Inceptions, we propose to decompose large-kernel depth-wise convolution into four parallel branches along channel dimension, i.e., small square kernel, two orthogonal band kernels, and an identity mapping. With this new Inception depthwise convolution, we build a series of networks, namely IncepitonNeXt, which not only enjoy high throughputs but also maintain competitive performance. For instance, InceptionNeXt-T achieves 1.6 × higher training throughputs than ConvNeX-T, as well as attains 0.2% top-1 accuracy improvement on ImageNet-1K. We antici-pate InceptionNeXt can serve as an economical baseline for future architecture design to reduce carbon footprint.

Even when using large multi-modal foundation models, few-shot learning is still challenging-if there is no proper inductive bias, it is nearly impossible to keep the nuanced class attributes while removing the visually prominent attributes that spuriously correlate with class labels. To this end, we find an inductive bias that the time-steps of a Diffusion Model (DM) can isolate the nuanced class attributes, i.e., as the forward diffusion adds noise to an image at each time-step, nuanced attributes are usually lost at an earlier time-step than the spurious attributes that are visually prominent. Building on this, we propose Time-step Few-shot (TiF) learner. We train class-specific low-rank adapters for a text-conditioned DM to make up for the lost attributes, such that images can be accurately reconstructed from their noisy ones given a prompt. Hence, at a small time-step, the adapter and prompt are essentially a parameterization of only the nuanced class attributes. For a test image, we can use the parameterization to only extract the nuanced class attributes for classification. TiF learner significantly outperforms OpenCLIP and its adapters on a variety of fine-grained and customized few-shot learning tasks. Codes are in https://github.com/yue-zhongqi/tif.

Chain-of-Thought (CoT) [2], [3] guides large language models (LLMs) to reason step-by-step, and can motivate their logical reasoning ability. While effective for logi-cal tasks, CoT is not conducive to creative problem-solving which often requires out-of-box thoughts and is crucial for innovation advancements. In this paper, we explore the Leap-of-Thought (LoT) abilities within LLMs - a non-sequential, creative paradigm involving strong associations and knowledge leaps. To this end, we study LLMs on the popular Oogiri game which needs participants to have good creativity and strong associative thinking for responding unexpectedly and humorously to the given image, text, or both, and thus is suitable for LoT study. Then to investi-gate LLMs' LoT ability in the Oogiri game, we first build a multimodal and multilingual Oogiri-GO dataset which contains over 130,000 samples from the Oogiri game, and observe the insufficient LoT ability or failures of most existing LLMs on the Oogiri game. Accordingly, we introduce a creative Leap-of-Thought (CLoT) paradigm to improve LLM's LoT ability. CLoT first formulates the Oogiri-GO dataset into LoT-oriented instruction tuning data to train pre-trained LLM for achieving certain LoT humor generation and discrimination abilities. Then CLoT designs an ex-plorative self-refinement that encourages the LLM to gener-ate more creative LoT data via exploring parallels between seemingly unrelated concepts and selects high-quality data to train itself for self-refinement. CLoT not only excels in humor generation in the Oogiri game as shown in Fig. 1 but also boosts creative abilities in various tasks like “cloud guessing game” and “divergent association task”. These findings advance our understanding and offer a pathway to improve LLMs' creative capacities for innovative applications across domains. The dataset, code, and models have been released online: https://zhongshsh.github.io/CLoT.

Second-order optimizers, maintaining a matrix termed a preconditioner, are superior to first-order optimizers in both theory and practice. The states forming the preconditioner and its inverse root restrict the maximum size of models trained by second-order optimizers. To address this, compressing 32-bit optimizer states to lower bitwidths has shown promise in reducing memory usage. However, current approaches only pertain to first-order optimizers. In this paper, we propose the first 4-bit second-order optimizers, exemplified by 4-bit Shampoo, maintaining performance similar to that of 32-bit ones. We show that quantizing the eigenvector matrix of the preconditioner in 4-bit Shampoo is remarkably better than quantizing the preconditioner itself both theoretically and experimentally. By rectifying the orthogonality of the quantized eigenvector matrix, we enhance the approximation of the preconditioner's eigenvector matrix, which also benefits the computation of its inverse 4-th root. Besides, we find that linear square quantization slightly outperforms dynamic tree quantization when quantizing second-order optimizer states. Evaluation on various networks for image classification and natural language modeling demonstrates that our 4-bit Shampoo achieves comparable performance to its 32-bit counterpart while being more memory-efficient.

Despite their success, unsupervised domain adaptation methods for semantic segmentation primarily focus on adaptation between image domains and do not utilize other abundant visual modalities like depth, infrared and event. This limitation hinders their performance and restricts their application in real-world multimodal scenarios. To address this issue, we propose Modality Adaptation with text-to-image Diffusion Models (MADM) for semantic segmentation task which utilizes text-to-image diffusion models pre-trained on extensive image-text pairs to enhance the model's cross-modality capabilities. Specifically, MADM comprises two key complementary components to tackle major challenges. First, due to the large modality gap, using one modal data to generate pseudo labels for another modality suffers from a significant drop in accuracy. To address this, MADM designs diffusion-based pseudo-label generation which adds latent noise to stabilize pseudo-labels and enhance label accuracy. Second, to overcome the limitations of latent low-resolution features in diffusion models, MADM introduces the label palette and latent regression which converts one-hot encoded labels into the RGB form by palette and regresses them in the latent space, thus ensuring the pre-trained decoder for up-sampling to obtain fine-grained features. Extensive experimental results demonstrate that MADM achieves state-of-the-art adaptation performance across various modality tasks, including images to depth, infrared, and event modalities. We open-source our code and models at https://github.com/XiaRho/MADM.

Recent 3D large reconstruction models (LRMs) can generate high-quality 3D content in sub-seconds by integrating multi-view diffusion models with scalable multi-view reconstructors. Current works further leverage 3D Gaussian Splatting as 3D representation for improved visual quality and rendering efficiency. However, we observe that existing Gaussian reconstruction models often suffer from multi-view inconsistency and blurred textures. We attribute this to the compromise of multi-view information propagation in favor of adopting powerful yet computationally intensive architectures (\eg, Transformers). To address this issue, we introduce MVGamba, a general and lightweight Gaussian reconstruction model featuring a multi-view Gaussian reconstructor based on the RNN-like State Space Model (SSM). Our Gaussian reconstructor propagates causal context containing multi-view information for cross-view self-refinement while generating a long sequence of Gaussians for fine-detail modeling with linear complexity. With off-the-shelf multi-view diffusion models integrated, MVGamba unifies 3D generation tasks from a single image, sparse images, or text prompts. Extensive experiments demonstrate that MVGamba outperforms state-of-the-art baselines in all 3D content generation scenarios with approximately only $0.1\times$ of the model size. The codes are available at \url{https://github.com/SkyworkAI/MVGamba}.

Question answering, asking, and assessment are three innate human traits crucial for understanding the world and acquiring knowledge. By enhancing these capabilities, humans can more effectively utilize data, leading to better comprehension and learning outcomes. However, current Multimodal Large Language Models (MLLMs) primarily focus on question answering, often neglecting the full potential of questioning and assessment skills. In this study, we introduce LOVA3, an innovative framework named ``Learning tO Visual Question Answering, Asking and Assessment,'' designed to equip MLLMs with these additional capabilities. Our approach involves the creation of two supplementary training tasks GenQA and EvalQA, aiming at fostering the skills of asking and assessing questions in the context of images. To develop the questioning ability, we compile a comprehensive set of multimodal foundational tasks. For assessment, we introduce a new benchmark called EvalQABench, comprising 64,000 training samples (split evenly between positive and negative samples) and 5,000 testing samples. We posit that enhancing MLLMs with the capabilities to answer, ask, and assess questions will enhance their multimodal comprehension, ultimately improving overall performance. To validate this hypothesis, we train MLLMs using the LOVA3 framework and evaluate them on a range of multimodal datasets and benchmarks. Our results demonstrate consistent performance gains, underscoring the critical role of these additional tasks in fostering comprehensive intelligence in MLLMs.

Training deep networks on large-scale datasets is computationally challenging. This work explores the problem of “how to accelerate adaptive gradient algorithms in a general manner", and proposes an effective Weight-decay-Integrated Nesterov acceleration (Win) to accelerate adaptive algorithms. Taking AdamW and Adam as examples, per iteration, we construct a dynamical loss that combines the vanilla training loss and a dynamic regularizer inspired by proximal point method, and respectively minimize the first- and second-order Taylor approximations of dynamical loss to update variable. This yields our Win acceleration that uses a conservative step and an aggressive step to update, and linearly combines these two updates for acceleration. Next, we extend Win into Win2 which uses multiple aggressive update steps for faster convergence. Then we apply Win and Win2 to the popular LAMB and SGD optimizers. Our transparent derivation could provide insights for other accelerated methods and their integration into adaptive algorithms. Besides, we theoretically justify the faster convergence of Win- and Win2-accelerated AdamW, Adam and LAMB to their non-accelerated counterparts. Experimental results demonstrates the faster convergence speed and superior performance of our Win- and Win2-accelerated AdamW, Adam, LAMB and SGD over their vanilla counterparts on vision classification and language modeling tasks.

Vision-Language Pre-Training (VLP) has demonstrated remarkable potential in aligning image and text pairs, paving the way for a wide range of cross-modal learning tasks. Nevertheless, we have observed that VLP models often fall short in terms of visual grounding and localization capabilities, which are crucial for many downstream tasks, such as visual reasoning. In response, we introduce a novel Position-guided Text Prompt (PTP) paradigm to bolster the visual grounding abilities of cross-modal models trained with VLP. In the VLP phase, PTP divides an image into N x N blocks and employs a widely-used object detector to identify objects within each block. PTP then reframes the visual grounding task as a fill-in-the-blank problem, encouraging the model to predict objects in given blocks or regress the blocks of a given object, exemplified by filling "[P]" or "[O]" in a PTP sentence such as "The block [P] has a [O]." This strategy enhances the visual grounding capabilities of VLP models, enabling them to better tackle various downstream tasks. Additionally, we integrate the seconda-order relationships between objects to further enhance the visual grounding capabilities of our proposed PTP paradigm. Incorporating PTP into several state-of-the-art VLP frameworks leads to consistently significant improvements across representative cross-modal learning model architectures and multiple benchmarks, such as zero-shot Flickr30 k Retrieval (+5.6 in average recall@1) for ViLT baseline, and COCO Captioning (+5.5 in CIDEr) for the state-of-the-art BLIP baseline. Furthermore, PTP attains comparable results with object-detector-based methods and a faster inference speed, as it discards its object detector during inference, unlike other approaches.

In deep learning, different kinds of deep networks typically need different optimizers, which have to be chosen after multiple trials, making the training process inefficient. To relieve this issue and consistently improve the model training speed across deep networks, we propose the ADAptive Nesterov momentum algorithm, Adan for short. Adan first reformulates the vanilla Nesterov acceleration to develop a new Nesterov momentum estimation (NME) method, which avoids the extra overhead of computing gradient at the extrapolation point. Then Adan adopts NME to estimate the gradient's first- and second-order moments in adaptive gradient algorithms for convergence acceleration. Besides, we prove that Adan finds an$\epsilon$-approximate first-order stationary point within$\mathcal {O}(\epsilon ^{-3.5})$stochastic gradient complexity on the non-convex stochastic problems (e.g., deep learning problems), matching the best-known lower bound. Extensive experimental results show that Adan consistently surpasses the corresponding SoTA optimizers on vision, language, and RL tasks and sets new SoTAs for many popular networks and frameworks, e.g., ResNet, ConvNext, ViT, Swin, MAE, DETR, GPT-2, Transformer-XL, and BERT. More surprisingly, Adan can use half of the training cost (epochs) of SoTA optimizers to achieve higher or comparable performance on ViT, GPT-2, MAE,etc, and also shows great tolerance to a large range of minibatch size, e.g., from 1 k to 32 k.

MetaFormer, the abstracted architecture of Transformer, has been found to play a significant role in achieving competitive performance. In this paper, we further explore the capacity of MetaFormer, again, by migrating our focus away from the token mixer design: we introduce several baseline models under MetaFormer using the most basic or common mixers, and demonstrate their gratifying performance. We summarize our observations as follows: (1) MetaFormer ensures solid lower bound of performance. By merely adopting identity mapping as the token mixer, the MetaFormer model, termed IdentityFormer, achieves [Formula: see text]80% accuracy on ImageNet-1 K. (2) MetaFormer works well with arbitrary token mixers. When specifying the token mixer as even a random matrix to mix tokens, the resulting model RandFormer yields an accuracy of [Formula: see text]81%, outperforming IdentityFormer. Rest assured of MetaFormer's results when new token mixers are adopted. (3) MetaFormer effortlessly offers state-of-the-art results. With just conventional token mixers dated back five years ago, the models instantiated from MetaFormer already beat state of the art. (a) ConvFormer outperforms ConvNeXt. Taking the common depthwise separable convolutions as the token mixer, the model termed ConvFormer, which can be regarded as pure CNNs, outperforms the strong CNN model ConvNeXt. (b) CAFormer sets new record on ImageNet-1 K. By simply applying depthwise separable convolutions as token mixer in the bottom stages and vanilla self-attention in the top stages, the resulting model CAFormer sets a new record on ImageNet-1 K: it achieves an accuracy of 85.5% at 224 ×224 resolution, under normal supervised training without external data or distillation. In our expedition to probe MetaFormer, we also find that a new activation, StarReLU, reduces 71% FLOPs of activation compared with commonly-used GELU yet achieves better performance. Specifically, StarReLU is a variant of Squared ReLU dedicated to alleviating distribution shift. We expect StarReLU to find great potential in MetaFormer- like models alongside other neural networks. Code and models are available at https://github.com/sail-sg/metaformer.

AdamW modifies Adam by adding a decoupled weight decay to decay network weights per training iteration. For adaptive algorithms, this decoupled weight decay does not affect specific optimization steps, and differs from the widely used$\ell _{2}$-regularizer which changes optimization steps via changing the first- and second-order gradient moments. Despite its great practical success, for AdamW, its convergence behavior and generalization improvement over Adam and$\ell _{2}$-regularized Adam ($\ell _{2}$-Adam) remain absent yet. To solve this issue, we prove the convergence of AdamW and justify its generalization advantages over Adam and$\ell _{2}$-Adam. Specifically, AdamW provably converges but minimizes a dynamically regularized loss that combines vanilla loss and a dynamical regularization induced by decoupled weight decay, thus yielding different behaviors with Adam and$\ell _{2}$-Adam. Moreover, on both general nonconvex problems and PŁ-conditioned problems, we establish stochastic gradient complexity of AdamW to find a stationary point. Such complexity is also applicable to Adam and$\ell _{2}$-Adam, and improves their previously known complexity, especially for over-parametrized networks. Besides, we prove that AdamW enjoys smaller generalization errors than Adam and$\ell _{2}$-Adam from the Bayesian posterior aspect. This result, for the first time, explicitly reveals the benefits of decoupled weight decay in AdamW. Experimental results validate our theory.

Recently, there has been increasing interest in the challenge of how to discriminatively vectorize graphs. To address this, we propose a method called Iterative Graph Self-Distillation (IGSD) which learns graph-level representation in an unsupervised manner through instance discrimination using a self-supervised contrastive learning approach. IGSD involves a teacher-student distillation process that uses graph diffusion augmentations and constructs the teacher model using an exponential moving average of the student model. The intuition behind IGSD is to predict the teacher network representation of the graph pairs under different augmented views. As a natural extension, we also apply IGSD to semi-supervised scenarios by jointly regularizing the network with both supervised and self-supervised contrastive loss. Finally, we show that fine-tuning the IGSD-trained models with self-training can further improve graph representation learning. Empirically, we achieve significant and consistent performance gain on various graph datasets in both unsupervised and semi-supervised settings, which well validates the superiority of IGSD.

Graph-level representations are critical in various real-world applications, such as predicting the properties of molecules. However, in practice, precise graph annotations are generally very expensive and time-consuming. To address this issue, graph contrastive learning constructs an instance discrimination task, which pulls together positive pairs (augmentation pairs of the same graph) and pushes away negative pairs (augmentation pairs of different graphs) for unsupervised representation learning. However, since for a query, its negatives are uniformly sampled from all graphs, existing methods suffer from the critical sampling bias issue, i.e., the negatives likely having the same semantic structure with the query, leading to performance degradation. To mitigate this sampling bias issue, in this article, we propose a prototypical graph contrastive learning (PGCL) approach. Specifically, PGCL models the underlying semantic structure of the graph data via clustering semantically similar graphs into the same group and simultaneously encourages the clustering consistency for different augmentations of the same graph. Then, given a query, it performs negative sampling via drawing the graphs from those clusters that differ from the cluster of query, which ensures the semantic difference between query and its negative samples. Moreover, for a query, PGCL further reweights its negative samples based on the distance between their prototypes (cluster centroids) and the query prototype such that those negatives having moderate prototype distance enjoy relatively large weights. This reweighting strategy is proven to be more effective than uniform sampling. Experimental results on various graph benchmarks testify the advantages of our PGCL over state-of-the-art methods. The code is publicly available at https://github.com/ha-lins/PGCL.

Vision-Language Pre-Training (VLP) has shown promising capabilities to align image and text pairs, facilitating a broad variety of cross-modal learning tasks. However, we observe that VLP models often lack the visual grounding/localization capability which is critical for many downstream tasks such as visual reasoning. In this work, we propose a novel Position-guided Text Prompt (PTP) paradigm to enhance the visual grounding ability of cross-modal models trained with VLP. Specifically, in the VLP phase, PTP divides the image into N x N blocks, and identifies the objects in each block through the widely used object detector in VLP. It then reformulates the visual grounding task into a fill-in-the-blank problem given a PTP by encouraging the model to predict the objects in the given blocks or regress the blocks of a given object, e.g. filling “[P]” or “[O]” in a PTP “The block [P] has a [O]”. This mechanism improves the visual grounding capability of VLP models and thus helps them better handle various downstream tasks. By introducing PTP into several state-of-the-art VLP frameworks, we observe consistently significant improvements across representative cross-modal learning model architectures and several benchmarks, e.g. zero-shot Flickr30K Retrieval (+4.8 in average recall@1) for ViLT [16] baseline, and COCO Captioning (+5.3 in CIDEr) for SOTA BLIP [19] baseline. Moreover, PTP achieves comparable results with object-detector based methods [8, 23, 45], and much faster inference speed since PTP discards its object detector for inference while the later cannot.

Despite its success in image synthesis, we observe that diffusion probabilistic models (DPMs) often lack contextual reasoning ability to learn the relations among object parts in an image, leading to a slow learning process. To solve this issue, we propose a Masked Diffusion Transformer (MDT) that introduces a mask latent modeling scheme to explicitly enhance the DPMs’ ability to contextual relation learning among object semantic parts in an image. During training, MDT operates in the latent space to mask certain tokens. Then, an asymmetric masking diffusion transformer is designed to predict masked tokens from unmasked ones while maintaining the diffusion generation process. Our MDT can reconstruct the full information of an image from its incomplete contextual input, thus enabling it to learn the associated relations among image tokens. Experimental results show that MDT achieves superior image synthesis performance, e.g., a new SOTA FID score in the ImageNet data set, and has about 3× faster learning speed than the previous SOTA DiT. The source code is released at https://github.com/sail-sg/MDT.

Existing methods of privacy-preserving action recognition (PPAR) mainly focus on frame-level (spatial) privacy removal through 2D CNNs. Unfortunately, they have two major drawbacks. First, they may compromise temporal dynamics in input videos, which are critical for accurate action recognition. Second, they are vulnerable to practical attacking scenarios where attackers probe for privacy from an entire video rather than individual frames. To address these issues, we propose a novel framework STPrivacy to perform video-level PPAR. For the first time, we introduce vision Transformers into PPAR by treating a video as a tubelet sequence, and accordingly design two complementary mechanisms, i.e., sparsification and anonymization, to remove privacy from a spatio-temporal perspective. In specific, our privacy sparsification mechanism applies adaptive token selection to abandon action-irrelevant tubelets. Then, our anonymization mechanism implicitly manipulates the remaining action-tubelets to erase privacy in the embedding space through adversarial learning. These mechanisms provide significant advantages in terms of privacy preservation for human eyes and action-privacy trade-off adjustment during deployment. We additionally contribute the first two large-scale PPAR benchmarks, VP-HMDB51 and VP-UCF101, to the community. Extensive evaluations on them, as well as two other tasks, validate the effectiveness and generalization capability of our framework.

Image editing plays a vital role in computer vision field, aiming to realistically manipulate images while ensuring seamless integration. It finds numerous applications across various fields. In this work, we present EditAnything, a novel approach that empowers users with unparalleled flexibility in editing and generating image content. EditAnything introduces an array of advanced features, including cross-image dragging (e.g., try-on), region-interactive editing, controllable layout generation, and virtual character replacement. By harnessing these capabilities, users can engage in interactive and flexible editing, giving captivating outcomes that uphold the integrity of the original image. With its diverse range of tools, EditAnything caters to a wide spectrum of editing needs, pushing the boundaries of image editing and unlocking exciting new possibilities. The source code is released at https://github.com/sail-sg/EditAnything.

In diffusion models, UNet is the most popular network backbone, since its long skip connects (LSCs) to connect distant network blocks can aggregate long-distant information and alleviate vanishing gradient. Unfortunately, UNet often suffers from unstable training in diffusion models which can be alleviated by scaling its LSC coefficients smaller. However, theoretical understandings of the instability of UNet in diffusion models and also the performance improvement of LSC scaling remain absent yet. To solve this issue, we theoretically show that the coefficients of LSCs in UNet have big effects on the stableness of the forward and backward propagation and robustness of UNet. Specifically, the hidden feature and gradient of UNet at any layer can oscillate and their oscillation ranges are actually large which explains the instability of UNet training. Moreover, UNet is also provably sensitive to perturbed input, and predicts an output distant from the desired output, yielding oscillatory loss and thus oscillatory gradient. Besides, we also observe the theoretical benefits of the LSC coefficient scaling of UNet in the stableness of hidden features and gradient and also robustness. Finally, inspired by our theory, we propose an effective coefficient scaling framework ScaleLong that scales the coefficients of LSC in UNet and better improve the training stability of UNet. Experimental results on CIFAR10, CelebA, ImageNet and COCO show that our methods are superior to stabilize training, and yield about 1.5x training acceleration on different diffusion models with UNet or UViT backbones.

We propose to perform video question answering (VideoQA) in a Contrastive manner via a Video Graph Transformer model (CoVGT). CoVGT's uniqueness and superiority are three-fold: 1) It proposes a dynamic graph transformer module which encodes video by explicitly capturing the visual objects, their relations and dynamics, for complex spatio-temporal reasoning. 2) It designs separate video and text transformers for contrastive learning between the video and text to perform QA, instead of multi-modal transformer for answer classification. Fine-grained video-text communication is done by additional cross-modal interaction modules. 3) It is optimized by the joint fully- and self-supervised contrastive objectives between the correct and incorrect answers, as well as the relevant and irrelevant questions respectively. With superior video encoding and QA solution, we show that CoVGT can achieve much better performances than previous arts on video reasoning tasks. Its performances even surpass those models that are pretrained with millions of external data. We further show that CoVGT can also benefit from cross-modal pretraining, yet with orders of magnitude smaller data. The results demonstrate the effectiveness and superiority of CoVGT, and additionally reveal its potential for more data-efficient pretraining.

Transformers have shown great potential in computer vision tasks. A common belief is their attention-based token mixer module contributes most to their competence. However, recent works show the attention-based module in transformers can be replaced by spatial MLPs and the resulted models still perform quite well. Based on this observation, we hypothesize that the general architecture of the transformers, instead of the specific token mixer module, is more essential to the model's performance. To verify this, we deliberately replace the attention module in transformers with an embarrassingly simple spatial pooling operator to conduct only basic token mixing. Surprisingly, we observe that the derived model, termed as PoolFormer, achieves competitive performance on multiple computer vision tasks. For example, on ImageNet-1K, PoolFormer achieves 82.1 % top-1 accuracy, surpassing well-tuned vision transformer/MLP-like baselines DeiT-B/ResMLP-B24 by 0.3%/1.1% accuracy with 35%/52% fewer parameters and 49%/61% fewer MACs. The effectiveness of Pool-Former verifies our hypothesis and urges us to initiate the concept of “MetaFormer”, a general architecture abstracted from transformers without specifying the token mixer. Based on the extensive experiments, we argue that MetaFormer is the key player in achieving superior results for recent transformer and MLP-like models on vision tasks. This work calls for more future research dedicated to improving MetaFormer instead of focusing on the token mixer modules. Additionally, our proposed PoolFormer could serve as a starting baseline for future MetaFormer architecture design.

Recent studies show that transformer has strong capability of building long-range dependencies, yet is incompetent in capturing high frequencies that predominantly convey local information. To tackle this issue, we present a novel and general-purpose $\textit{Inception Transformer}$, or $\textit{iFormer}$ for short, that effectively learns comprehensive features with both high- and low-frequency information in visual data. Specifically, we design an Inception mixer to explicitly graft the advantages of convolution and max-pooling for capturing the high-frequency information to transformers. Different from recent hybrid frameworks, the Inception mixer brings greater efficiency through a channel splitting mechanism to adopt parallel convolution/max-pooling path and self-attention path as high- and low-frequency mixers, while having the flexibility to model discriminative information scattered within a wide frequency range. Considering that bottom layers play more roles in capturing high-frequency details while top layers more in modeling low-frequency global information, we further introduce a frequency ramp structure, i.e., gradually decreasing the dimensions fed to the high-frequency mixer and increasing those to the low-frequency mixer, which can effectively trade-off high- and low-frequency components across different layers. We benchmark the iFormer on a series of vision tasks, and showcase that it achieves impressive performance on image classification, COCO detection and ADE20K segmentation. For example, our iFormer-S hits the top-1 accuracy of 83.4% on ImageNet-1K, much higher than DeiT-S by 3.6%, and even slightly better than much bigger model Swin-B (83.3%) with only 1/4 parameters and 1/3 FLOPs. Code and models are released at https://github.com/sail-sg/iFormer.

Despite the success of stochastic variance-reduced gradient (SVRG) algorithms in solving large-scale problems, their stochastic gradient complexity often scales linearly with data size and is expensive for huge data. Accordingly, we propose a hybrid stochastic-deterministic minibatch proximal gradient (HSDMPG) algorithm for strongly convex problems with linear prediction structure, e.g., least squares and logistic/softmax regression.HSDMPGenjoys improved computational complexity that is data-size-independent for large-scale problems. It iteratively samples an evolving minibatch of individual losses to estimate the original problem, and can efficiently minimize the sampled subproblems. For strongly convex loss of$n$components,HSDMPGattains an$\epsilon$-optimization-error within$\mathcal {O} \left(\kappa \log ^{\zeta +1}\left(\frac{1}{\epsilon }\right)\frac{1}{\epsilon }\bigwedge n\log ^{\zeta }\left(\frac{1}{\epsilon }\right)\right)$stochastic gradient evaluations, where$\kappa$is condition number,$\zeta =1$for quadratic loss and$\zeta =2$for generic loss. For large-scale problems, our complexity outperforms those of SVRG-type algorithms with/without dependence on data size. Particularly, when$\epsilon =\mathcal {O}(1/\sqrt{n})$which matches the intrinsic excess error of a learning model and is sufficient for generalization, our complexity for quadratic and generic losses is respectively$\mathcal {O} (n^{0.5}\log ^{2}(n))$and$\mathcal {O} (n^{0.5}\log ^{3}(n))$, which for the first time achieves optimal generalization in less than a single pass over data. Besides, we extendHSDMPGto online strongly convex problems and prove its higher efficiency over the prior algorithms. Numerical results demonstrate the computational advantages ofHSDMPG.

Recently, stochastic hard thresholding (HT) optimization methods [e.g., stochastic variance reduced gradient hard thresholding (SVRGHT)] are becoming more attractive for solving large-scale sparsity/rank-constrained problems. However, they have much higher HT oracle complexities, especially for high-dimensional data or large-scale matrices. To address this issue and inspired by the well-known Gradient Support Pursuit (GraSP) method, this article proposes a new Relaxed Gradient Support Pursuit (RGraSP) framework. Unlike GraSP, RGraSP only requires to yield an approximation solution at each iteration. Based on the property of RGraSP, we also present an efficient stochastic variance reduction-gradient support pursuit algorithm and its fast version (called stochastic variance reduced gradient support pursuit (SVRGSP+). We prove that the gradient oracle complexity of both our algorithms is two times less than that of SVRGHT. In particular, their HT complexity is about$\kappa _{\widehat {s}}$times less than that of SVRGHT, where$\kappa _{\widehat {s}}$is the restricted condition number. Moreover, we prove that our algorithms enjoy fast linear convergence to an approximately global optimum, and also present an asynchronous parallel variant to deal with very high-dimensional and sparse data. Experimental results on both synthetic and real-world datasets show that our algorithms yield superior results than the state-of-the-art gradient HT methods.

Human doctors with well-structured medical knowledge can diagnose a disease merely via a few conversations with patients about symptoms. In contrast, existing knowledge-grounded dialogue systems often require a large number of dialogue instances to learn as they fail to capture the correlations between different diseases and neglect the diagnostic experience shared among them. To address this issue, we propose a more natural and practical paradigm, i.e., low-resource medical dialogue generation, which can transfer the diagnostic experience from source diseases to target ones with a handful of data for adaptation. It is capitalized on a commonsense knowledge graph to characterize the prior disease-symptom relations. Besides, we develop a Graph-Evolving Meta-Learning (GEML) framework that learns to evolve the commonsense graph for reasoning disease-symptom correlations in a new disease, which effectively alleviates the needs of a large number of dialogues. More importantly, by dynamically evolving disease-symptom graphs, GEML also well addresses the real-world challenges that the disease-symptom correlations of each disease may vary or evolve along with more diagnostic cases. Extensive experiment results on the CMDD dataset and our newly-collected Chunyu dataset testify the superiority of our approach over state-of-the-art approaches. Besides, our GEML can generate an enriched dialogue-sensitive knowledge graph in an online manner, which could benefit other tasks grounded on knowledge graph.

Low-resource automatic speech recognition (ASR) is challenging, as the low-resource target language data cannot well train an ASR model. To solve this issue, meta-learning formulates ASR for each source language into many small ASR tasks and meta-learns a model initialization on all tasks from different source languages to access fast adaptation on unseen target languages. However, for different source languages, the quantity and difficulty vary greatly because of their different data scales and diverse phonological systems, which leads to task-quantity and task-difficulty imbalance issues and thus a failure of multilingual meta-learning ASR (MML-ASR). In this work, we solve this problem by developing a novel adversarial meta sampling (AMS) approach to improve MML-ASR. When sampling tasks in MML-ASR, AMS adaptively determines the task sampling probability for each source language. Specifically, for each source language, if the query loss is large, it means that its tasks are not well sampled to train ASR model in terms of its quantity and difficulty and thus should be sampled more frequently for extra learning. Inspired by this fact, we feed the historical task query loss of all source language domain into a network to learn a task sampling policy for adversarially increasing the current query loss of MML-ASR. Thus, the learnt task sampling policy can master the learning situation of each language and thus predicts good task sampling probability for each language for more effective learning. Finally, experiment results on two multilingual datasets show significant performance improvement when applying our AMS on MML-ASR, and also demonstrate the applicability of AMS to other low-resource speech tasks and transfer learning ASR approaches.

Meta-learning aims to perform fast adaptation on a new task through learning a “prior” from multiple existing tasks. A common practice in meta-learning is to perform a train-validation split (\emph{train-val method}) where the prior adapts to the task on one split of the data, and the resulting predictor is evaluated on another split. Despite its prevalence, the importance of the train-validation split is not well understood either in theory or in practice, particularly in comparison to the more direct \emph{train-train method}, which uses all the per-task data for both training and evaluation. We provide a detailed theoretical study on whether and when the train-validation split is helpful in the linear centroid meta-learning problem. In the agnostic case, we show that the expected loss of the train-val method is minimized at the optimal prior for meta testing, and this is not the case for the train-train method in general without structural assumptions on the data. In contrast, in the realizable case where the data are generated from linear models, we show that both the train-val and train-train losses are minimized at the optimal prior in expectation. Further, perhaps surprisingly, our main result shows that the train-train method achieves a \emph{strictly better} excess loss in this realizable case, even when the regularization parameter and split ratio are optimally tuned for both methods. Our results highlight that sample splitting may not always be preferable, especially when the data is realizable by the model. We validate our theories by experimentally showing that the train-train method can indeed outperform the train-val method, on both simulations and real meta-learning tasks.

For an image query, unsupervised contrastive learning labels crops of the same image as positives, and other image crops as negatives. Although intuitive, such a native label assignment strategy cannot reveal the underlying semantic similarity between a query and its positives and negatives, and impairs performance, since some negatives are semantically similar to the query or even share the same semantic class as the query. In this work, we first prove that for contrastive learning, inaccurate label assignment heavily impairs its generalization for semantic instance discrimination, while accurate labels benefit its generalization. Inspired by this theory, we propose a novel self-labeling refinement approach for contrastive learning. It improves the label quality via two complementary modules: (i) self-labeling refinery (SLR) to generate accurate labels and (ii) momentum mixup (MM) to enhance similarity between query and its positive. SLR uses a positive of a query to estimate semantic similarity between a query and its positive and negatives, and combines estimated similarity with vanilla label assignment in contrastive learning to iteratively generate more accurate and informative soft labels. We theoretically show that our SLR can exactly recover the true semantic labels of label-corrupted data, and supervises networks to achieve zero prediction error on classification tasks. MM randomly combines queries and positives to increase semantic similarity between the generated virtual queries and their positives so as to improves label accuracy. Experimental results on CIFAR10, ImageNet, VOC and COCO show the effectiveness of our method.

To train networks, lookahead algorithm~\cite{zhang2019lookahead} updates its fast weights $k$ times via an inner-loop optimizer before updating its slow weights once by using the latest fast weights. Any optimizer, e.g. SGD, can serve as the inner-loop optimizer, and the derived lookahead generally enjoys remarkable test performance improvement over the vanilla optimizer. But theoretical understandings on the test performance improvement of lookahead remain absent yet. To solve this issue, we theoretically justify the advantages of lookahead in terms of the excess risk error which measures the test performance. Specifically, we prove that lookahead using SGD as its inner-loop optimizer can better balance the optimization error and generalization error to achieve smaller excess risk error than vanilla SGD on (strongly) convex problems and nonconvex problems with Polyak-{\L}ojasiewicz condition which has been observed/proved in neural networks. Moreover, we show the stagewise optimization strategy~\cite{barshan2015stage} which decays learning rate several times during training can also benefit lookahead in improving its optimization and generalization errors on strongly convex problems. Finally, we propose a stagewise locally-regularized lookahead (SLRLA) algorithm which sums up the vanilla objective and a local regularizer to minimize at each stage and provably enjoys optimization and generalization improvement over the conventional (stagewise) lookahead. Experimental results on CIFAR10/100 and ImageNet testify its advantages. Codes is available at \url{https://github.com/sail-sg/SLRLA-optimizer}.

Few-shot learning ability is heavily desired for machine intelligence. By meta-learning a model initialization from training tasks with fast adaptation ability to new tasks, model-agnostic meta-learning (MAML) has achieved remarkable success in a number of few-shot learning applications. However, theoretical understandings on the learning ability of MAML remain absent yet, hindering developing new and more advanced meta learning methods in a principled way. In this work, we solve this problem by theoretically justifying the fast adaptation capability of MAML when applied to new tasks. Specifically, we prove that the learnt meta-initialization can benefit the fast adaptation to new tasks with only a few steps of gradient descent. This result explicitly reveals the benefits of the unique designs in MAML. Then we propose a theory-inspired task similarity aware MAML which clusters tasks into multiple groups according to the estimated optimal model parameters and learns group-specific initializations. The proposed method improves upon MAML by speeding up the adaptation and giving stronger few-shot learning ability. Experimental results on the few-shot classification tasks testify its advantages.

Multi-way or tensor data analysis has attracted increasing attention recently, with many important applications in practice. This article develops a tensor low-rank representation (TLRR) method, which is the first approach that can exactly recover the clean data of intrinsic low-rank structure and accurately cluster them as well, with provable performance guarantees. In particular, for tensor data with arbitrary sparse corruptions, TLRR can exactly recover the clean data under mild conditions; meanwhile TLRR can exactly verify their true origin tensor subspaces and hence cluster them accurately. TLRR objective function can be optimized via efficient convex programing with convergence guarantees. Besides, we provide two simple yet effective dictionary construction methods, the simple TLRR (S-TLRR) and robust TLRR (R-TLRR), to handle slightly and severely corrupted data respectively. Experimental results on two computer vision data analysis tasks, image/video recovery and face clustering, clearly demonstrate the superior performance, efficiency and robustness of our developed method over state-of-the-arts including the popular LRR and SSC methods.

First-order non-convex Riemannian optimization algorithms have gained recent popularity in structured machine learning problems including principal component analysis and low-rank matrix completion. The current paper presents an efficient Riemannian Stochastic Path Integrated Differential EstimatoR (R-SPIDER) algorithm to solve the finite-sum and online Riemannian non-convex minimization problems. At the core of R-SPIDER is a recursive semi-stochastic gradient estimator that can accurately estimate Riemannian gradient under not only exponential mapping and parallel transport, but also general retraction and vector transport operations. Compared with prior Riemannian algorithms, such a recursive gradient estimation mechanism endows R-SPIDER with lower computational cost in first-order oracle complexity. Specifically, for finite-sum problems with n components, R-SPIDER is proved to converge to an ϵ-approximate stationary point within [Formula: see text] stochastic gradient evaluations, beating the best-known complexity [Formula: see text]; for online optimization, R-SPIDER is shown to converge with [Formula: see text] complexity which is, to the best of our knowledge, the first non-asymptotic result for online Riemannian optimization. For the special case of gradient dominated functions, we further develop a variant of R-SPIDER with improved linear rate of convergence. Extensive experimental results demonstrate the advantage of the proposed algorithms over the state-of-the-art Riemannian non-convex optimization methods.

Stochastic variance-reduced gradient (SVRG) algorithms have been shown to work favorably in solving large-scale learning problems. Despite the remarkable success, the stochastic gradient complexity of SVRG-type algorithms usually scales linearly with data size and thus could still be expensive for huge data. To address this deficiency, we propose a hybrid stochastic-deterministic minibatch proximal gradient (\HSDAN) algorithm for strongly-convex problems that enjoys provably improved data-size-independent complexity guarantees. More precisely, for quadratic loss $F(\wm)$ of $n$ components, we prove that \HSDAN can attain an $\epsilon$-optimization-error $\EE[F(\wm)-F(\wms)] \leq \epsilon$ within $\mathcal{O}\Big(\frac{\kappa^{1.5}\epsilon^{0.75} \log^{1.5}(\frac{1}{\epsilon}) + 1}{\epsilon} \wedge \Big(\kappa \sqrt{n} \log^{1.5}\big(\frac{1}{\epsilon}\big) + n \log \big(\frac{1}{\epsilon}\big) \Big) \Big)$ stochastic gradient evaluations, where $\kappa$ is condition number. For generic strongly convex loss functions, we prove a nearly identical complexity bound though at the cost of slightly increased logarithmic factors. For large-scale learning problems, our complexity bounds are superior to those of the prior state-of-the-art SVRG algorithms with or without dependence on data size. Particularly, in the case of $\epsilon\!=\!\mathcal{O}\big(1/\sqrt{n}\big)$ which is at the order of intrinsic excess error bound of a learning model and thus sufficient for generalization, the stochastic gradient complexity bounds of \HSDAN for quadratic and generic loss functions are respectively $\mathcal{O} (n^{0.875}\log^{1.5}(n))$ and $\mathcal{O} (n^{0.875}\log^{2.25}(n))$, which to our best knowledge, for the first time achieve optimal generalization in less than a single pass over data. Extensive numerical results demonstrate the computational advantages of our algorithm over the prior ones.

Despite success on a wide range of problems related to vision, generative adversarial networks (GANs) often suffer from inferior performance due to unstable training, especially for text generation. To solve this issue, we propose a new variational GAN training framework which enjoys superior training stability. Our approach is inspired by a connection of GANs and reinforcement learning under a variational perspective. The connection leads to (1) probability ratio clipping that regularizes generator training to prevent excessively large updates, and (2) a sample re-weighting mechanism that improves discriminator training by downplaying bad-quality fake samples. Moreover, our variational GAN framework can provably overcome the training issue in many GANs that an optimal discriminator cannot provide any informative gradient to training generator. By plugging the training approach in diverse state-of-the-art GAN architectures, we obtain significantly improved performance over a range of tasks, including text generation, text style transfer, and image generation.

It is not clear yet why ADAM-alike adaptive gradient algorithms suffer from worse generalization performance than SGD despite their faster training speed. This work aims to provide understandings on this generalization gap by analyzing their local convergence behaviors. Specifically, we observe the heavy tails of gradient noise in these algorithms. This motivates us to analyze these algorithms through their Levy-driven stochastic differential equations (SDEs) because of the similar convergence behaviors of an algorithm and its SDE. Then we establish the escaping time of these SDEs from a local basin. The result shows that (1) the escaping time of both SGD and ADAM~depends on the Radon measure of the basin positively and the heaviness of gradient noise negatively; (2) for the same basin, SGD enjoys smaller escaping time than ADAM, mainly because (a) the geometry adaptation in ADAM~via adaptively scaling each gradient coordinate well diminishes the anisotropic structure in gradient noise and results in larger Radon measure of a basin; (b) the exponential gradient average in ADAM~smooths its gradient and leads to lighter gradient noise tails than SGD. So SGD is more locally unstable than ADAM~at sharp minima defined as the minima whose local basins have small Radon measure, and can better escape from them to flatter ones with larger Radon measure. As flat minima here which often refer to the minima at flat or asymmetric basins/valleys often generalize better than sharp ones~\cite{keskar2016large,he2019asymmetric}, our result explains the better generalization performance of SGD over ADAM. Finally, experimental results confirm our heavy-tailed gradient noise assumption and theoretical affirmation.

Despite its high search efficiency, differential architecture search (DARTS) often selects network architectures with dominated skip connections which lead to performance degradation. However, theoretical understandings on this issue remain absent, hindering the development of more advanced methods in a principled way. In this work, we solve this problem by theoretically analyzing the effects of various types of operations, e.g. convolution, skip connection and zero operation, to the network optimization. We prove that the architectures with more skip connections can converge faster than the other candidates, and thus are selected by DARTS. This result, for the first time, theoretically and explicitly reveals the impact of skip connections to fast network optimization and its competitive advantage over other types of operations in DARTS. Then we propose a theory-inspired path-regularized DARTS that consists of two key modules: (i) a differential group-structured sparse binary gate introduced for each operation to avoid unfair competition among operations, and (ii) a path-depth-wise regularization used to incite search exploration for deep architectures that often converge slower than shallow ones as shown in our theory and are not well explored during search. Experimental results on image classification tasks validate its advantages. Codes and models will be released.

SPIDER (Stochastic Path Integrated Differential EstimatoR) is an efficient gradient estimation technique developed for non-convex stochastic optimization. Although having been shown to attain nearly optimal computational complexity bounds, the SPIDER-type methods are limited to linear metric spaces. In this paper, we introduce the Riemannian SPIDER (R-SPIDER) method as a novel nonlinear-metric extension of SPIDER for efficient non-convex optimization on Riemannian manifolds. We prove that for finite-sum problems with $n$ components, R-SPIDER converges to an $\epsilon$-accuracy stationary point within $\mathcal{O}\big(\min\big(n+\frac{\sqrt{n}}{\epsilon^2},\frac{1}{\epsilon^3}\big)\big)$ stochastic gradient evaluations, which is sharper in magnitude than the prior Riemannian first-order methods. For online optimization, R-SPIDER is shown to converge with $\mathcal{O}\big(\frac{1}{\epsilon^3}\big)$ complexity which is, to the best of our knowledge, the first non-asymptotic result for online Riemannian optimization. Especially, for gradient dominated functions, we further develop a variant of R-SPIDER and prove its linear convergence rate. Numerical results demonstrate the computational efficiency of the proposed methods.

Majorization-Minimization (MM) algorithms optimize an objective function by iteratively minimizing its majorizing surrogate and offer attractively fast convergence rate for convex problems. However, their convergence behaviors for non-convex problems remain unclear. In this paper, we propose a novel MM surrogate function from strictly upper bounding the objective to bounding the objective in expectation. With this generalized surrogate conception, we develop a new optimization algorithm, termed SPI-MM, that leverages the recent proposed SPIDER for more efficient non-convex optimization. We prove that for finite-sum problems, the SPI-MM algorithm converges to an stationary point within deterministic and lower stochastic gradient complexity. To our best knowledge, this work gives the first non-asymptotic convergence analysis for MM-alike algorithms in general non-convex optimization. Extensive empirical studies on non-convex logistic regression and sparse PCA demonstrate the advantageous efficiency of the proposed algorithm and validate our theoretical results.

We address the problem of meta-learning which learns a prior over hypothesis from a sample of meta-training tasks for fast adaptation on meta-testing tasks. A particularly simple yet successful paradigm for this research is model-agnostic meta-learning (MAML). Implementation and analysis of MAML, however, can be tricky; first-order approximation is usually adopted to avoid directly computing Hessian matrix but as a result the convergence and generalization guarantees remain largely mysterious for MAML. To remedy this deficiency, in this paper we propose a minibatch proximal update based meta-learning approach for learning to efficient hypothesis transfer. The principle is to learn a prior hypothesis shared across tasks such that the minibatch risk minimization biased regularized by this prior can quickly converge to the optimal hypothesis in each training task. The prior hypothesis training model can be efficiently optimized via SGD with provable convergence guarantees for both convex and non-convex problems. Moreover, we theoretically justify the benefit of the learnt prior hypothesis for fast adaptation to new few-shot learning tasks via minibatch proximal update. Experimental results on several few-shot regression and classification tasks demonstrate the advantages of our method over state-of-the-arts.

Multi-task learning is popular in machine learning and computer vision. In multitask learning, properly modeling task relations is important for boosting the performance of jointly learned tasks. Task covariance modeling has been successfully used to model the relations of tasks but is limited to homogeneous multi-task learning. In this paper, we propose a feature based task relation modeling approach, suitable for both homogeneous and heterogeneous multi-task learning. First, we propose a new metric to quantify the relations between tasks. Based on the quantitative metric, we then develop the task relation layer, which can be combined with any deep learning architecture to form task relation networks to fully exploit the relations of different tasks in an online fashion. Benefiting from the task relation layer, the task relation networks can better leverage the mutual information from the data. We demonstrate our proposed task relation networks are effective in improving the performance in both homogeneous and heterogeneous multi-task learning settings through extensive experiments on computer vision tasks.

Most existing subspace clustering methods hinge on self-expression of handcrafted representations and are unaware of potential clustering errors. Thus they perform unsatisfactorily on real data with complex underlying subspaces. To solve this issue, we propose a novel deep adversarial subspace clustering (DASC) model, which learns more favorable sample representations by deep learning for subspace clustering, and more importantly introduces adversarial learning to supervise sample representation learning and subspace clustering. Specifically, DASC consists of a subspace clustering generator and a quality-verifying discriminator, which learn against each other. The generator produces subspace estimation and sample clustering. The discriminator evaluates current clustering performance by inspecting whether the re-sampled data from estimated subspaces have consistent subspace properties, and supervises the generator to progressively improve subspace clustering. Experimental results on the handwritten recognition, face and object clustering tasks demonstrate the advantages of DASC over shallow and few deep subspace clustering models. Moreover, to our best knowledge, this is the first successful application of GAN-alike model for unsupervised subspace clustering, which also paves the way for deep learning to solve other unsupervised learning problems.

This work aims to provide understandings on the remarkable success of deep convolutional neural networks (CNNs) by theoretically analyzing their generalization performance and establishing optimization guarantees for gradient descent based training algorithms. Specifically, for a CNN model consisting of $l$ convolutional layers and one fully connected layer, we prove that its generalization error is bounded by $\mathcal{O}(\sqrt{\theta\widetilde{\varrho}/n})$ where $\theta$ denotes freedom degree of the network parameters and $\widetilde{\varrho}=\mathcal{O}(\log(\prod_{i=1}^{l}b_{i} (k_{i}-s_{i}+1)/p)+\log(b_{l+1}))$ encapsulates architecture parameters including the kernel size $k_{i}$, stride $s_{i}$, pooling size $p$ and parameter magnitude $b_{i}$. To our best knowledge, this is the first generalization bound that only depends on $\mathcal{O}(\log(\prod_{i=1}^{l+1}b_{i}))$, tighter than existing ones that all involve an exponential term like $\mathcal{O}(\prod_{i=1}^{l+1}b_{i})$. Besides, we prove that for an arbitrary gradient descent algorithm, the computed approximate stationary point by minimizing empirical risk is also an approximate stationary point to the population risk. This well explains why gradient descent training algorithms usually perform sufficiently well in practice. Furthermore, we prove the one-to-one correspondence and convergence guarantees for the non-degenerate stationary points between the empirical and population risks. It implies that the computed local minimum for the empirical risk is also close to a local minimum for the population risk, thus ensuring that the optimized CNN model well generalizes to new data.

As an incremental-gradient algorithm, the hybrid stochastic gradient descent (HSGD) enjoys merits of both stochastic and full gradient methods for finite-sum minimization problem. However, the existing rate-of-convergence analysis for HSGD is made under with-replacement sampling (WRS) and is restricted to convex problems. It is not clear whether HSGD still carries these advantages under the common practice of without-replacement sampling (WoRS) for non-convex problems. In this paper, we affirmatively answer this open question by showing that under WoRS and for both convex and non-convex problems, it is still possible for HSGD (with constant step-size) to match full gradient descent in rate of convergence, while maintaining comparable sample-size-independent incremental first-order oracle complexity to stochastic gradient descent. For a special class of finite-sum problems with linear prediction models, our convergence results can be further improved in some cases. Extensive numerical results confirm our theoretical affirmation and demonstrate the favorable efficiency of WoRS-based HSGD.

Stochastic gradient hard thresholding methods have recently been shown to work favorably in solving large-scale empirical risk minimization problems under sparsity or rank constraint. Despite the improved iteration complexity over full gradient methods, the gradient evaluation and hard thresholding complexity of the existing stochastic algorithms usually scales linearly with data size, which could still be expensive when data is huge and the hard thresholding step could be as expensive as singular value decomposition in rank-constrained problems. To address these deficiencies, we propose an efficient hybrid stochastic gradient hard thresholding (HSG-HT) method that can be provably shown to have sample-size-independent gradient evaluation and hard thresholding complexity bounds. Specifically, we prove that the stochastic gradient evaluation complexity of HSG-HT scales linearly with inverse of sub-optimality and its hard thresholding complexity scales logarithmically. By applying the heavy ball acceleration technique, we further propose an accelerated variant of HSG-HT which can be shown to have improved factor dependence on restricted condition number. Numerical results confirm our theoretical affirmation and demonstrate the computational efficiency of the proposed methods.

Recently, a tensor nuclear norm (TNN) based method was proposed to solve the tensor completion problem, which has achieved state-of-the-art performance on image and video inpainting tasks. However, it requires computing tensor singular value decomposition (t-SVD), which costs much computation and thus cannot efficiently handle tensor data, due to its natural large scale. Motivated by TNN, we propose a novel low-rank tensor factorization method for efficiently solving the 3-way tensor completion problem. Our method preserves the low-rank structure of a tensor by factorizing it into the product of two tensors of smaller sizes. In the optimization process, our method only needs to update two smaller tensors, which can be more efficiently conducted than computing t-SVD. Furthermore, we prove that the proposed alternating minimization algorithm can converge to a Karush-Kuhn-Tucker point. Experimental results on the synthetic data recovery, image and video inpainting tasks clearly demonstrate the superior performance and efficiency of our developed method over state-of-the-arts including the TNN and matricization methods.

Low-rank tensor analysis is important for various real applications in computer vision. However, existing methods focus on recovering a low-rank tensor contaminated by Gaussian or gross sparse noise and hence cannot effectively handle outliers that are common in practical tensor data. To solve this issue, we propose an outlier-robust tensor principle component analysis (OR-TPCA) method for simultaneous low-rank tensor recovery and outlier detection. For intrinsically low-rank tensor observations with arbitrary outlier corruption, OR-TPCA is the first method that has provable performance guarantee for exactly recovering the tensor subspace and detecting outliers under mild conditions. Since tensor data are naturally high-dimensional and multi-way, we further develop a fast randomized algorithm that requires small sampling size yet can substantially accelerate OR-TPCA without performance drop. Experimental results on four tasks: outlier detection, clustering, semi-supervised and supervised learning, clearly demonstrate the advantages of our method.

Most supervised dictionary learning methods optimize the combinations of reconstruction error, sparsity prior, and discriminative terms. Thus, the learnt dictionaries may not be optimal for recognition tasks. Also, the sparse codes learning models in the training and the testing phases are inconsistent. Besides, without utilizing the intrinsic data structure, many dictionary learning methods only employ the$\ell _{0}$or$\ell _{1}$norm to encode each datum independently, limiting the performance of the learnt dictionaries. We present a novel bilevel model-based discriminative dictionary learning method for recognition tasks. The upper level directly minimizes the classification error, while the lower level uses the sparsity term and the Laplacian term to characterize the intrinsic data structure. The lower level is subordinate to the upper level. Therefore, our model achieves an overall optimality for recognition in that the learnt dictionary is directly tailored for recognition. Moreover, the sparse codes learning models in the training and the testing phases can be the same. We further propose a novel method to solve our bilevel optimization problem. It first replaces the lower level with its Karush-Kuhn–Tucker conditions and then applies the alternating direction method of multipliers to solve the equivalent problem. Extensive experiments demonstrate the effectiveness and robustness of our method.

Feature learning plays a central role in pattern recognition. In recent years, many representation-based feature learning methods have been proposed and have achieved great success in many applications. However, these methods perform feature learning and subsequent classification in two separate steps, which may not be optimal for recognition tasks. In this paper, we present a supervised low-rank-based approach for learning discriminative features. By integrating latent low-rank representation (LatLRR) with a ridge regression-based classifier, our approach combines feature learning with classification, so that the regulated classification error is minimized. In this way, the extracted features are more discriminative for the recognition tasks. Our approach benefits from a recent discovery on the closed-form solutions to noiseless LatLRR. When there is noise, a robust Principal Component Analysis (PCA)-based denoising step can be added as preprocessing. When the scale of a problem is large, we utilize a fast randomized algorithm to speed up the computation of robust PCA. Extensive experimental results demonstrate the effectiveness and robustness of our method.