Prior efforts in light-weight model development mainly centered on CNN and Transformer-based designs yet faced persistent challenges. CNNs adept at local feature extraction compromise resolution while Transformers offer global reach but escalate computational demands O(N^2). This ongoing trade-off between accuracy and efficiency remains a significant hurdle. Recently, state space models (SSMs), such as Mamba, have shown outstanding performance and competitiveness in various tasks such as language modeling and computer vision, while reducing the time complexity of global information extraction to O(N). Inspired by this, this work proposes to explore the potential of visual state space models in light-weight model design and introduce a novel efficient model variant dubbed EfficientVMamba. Concretely, our EfficientVMamba integrates a atrous-based selective scan approach by efficient skip sampling, constituting building blocks designed to harness both global and local representational features. Additionally, we investigate the integration between SSM blocks and convolutions, and introduce an efficient visual state space block combined with an additional convolution branch, which further elevate the model performance. Experimental results show that, EfficientVMamba scales down the computational complexity while yields competitive results across a variety of vision tasks. For example, our EfficientVMamba-S with 1.3G FLOPs improves Vim-Ti with 1.5G FLOPs by a large margin of 5.6% accuracy on ImageNet.

Visual encoders are fundamental components in vision-language models (VLMs), each showcasing unique strengths derived from various pre-trained visual foundation models. To leverage the various capabilities of these encoders, recent studies incorporate multiple encoders within a single VLM, leading to a considerable increase in computational cost. In this paper, we present Mixture-of-Visual-Encoder Knowledge Distillation (MoVEKD), a novel framework that distills the unique proficiencies of multiple vision encoders into a single, efficient encoder model. Specifically, to mitigate conflicts and retain the unique characteristics of each teacher encoder, we employ low-rank adaptation (LoRA) and mixture-of-experts (MoEs) to selectively activate specialized knowledge based on input features, enhancing both adaptability and efficiency. To regularize the KD process and enhance performance, we propose an attention-based distillation strategy that adaptively weighs the different encoders and emphasizes valuable visual tokens, reducing the burden of replicating comprehensive but distinct features from multiple teachers. Comprehensive experiments on popular VLMs, such as LLaVA and LLaVA-NeXT, validate the effectiveness of our method. Our code is available at: https://github.com/hey-cjj/MoVE-KD.

Masked autoencoders (MAEs) represent a prominent self-supervised learning paradigm in computer vision. Despite their empirical success, the underlying mechanisms of MAEs remain insufficiently understood. Recent studies have attempted to elucidate the functioning of MAEs through contrastive learning and feature representation analysis, yet these approaches often provide only implicit insights. In this paper, we propose a new perspective for understanding MAEs by leveraging the information bottleneck principle in information theory. Our theoretical analyses reveal that optimizing the latent features to balance relevant and irrelevant information is key to improving MAE performance. Building upon our proofs, we introduce MI-MAE, a novel method that optimizes MAEs through mutual information maximization and minimization. By enhancing latent features to retain maximal relevant information between them and the output, and minimizing irrelevant information between them and the input, our approach achieves better performance. Extensive experiments on standard benchmarks show that MI-MAE significantly outperforms MAE models in tasks such as image classification, object detection, and semantic segmentation. Our findings validate the theoretical framework and highlight the practical advantages of applying the information bottleneck principle to MAEs, offering deeper insights for developing more powerful self-supervised learning models.

Diffusion models have demonstrated remarkable efficacy in various generative tasks with the predictive prowess of denoising model. Currently, diffusion models employ a uniform denoising model across all timesteps. However, the inherent variations in data distributions at different timesteps lead to conflicts during training, constraining the potential of diffusion models. To address this challenge, we propose a novel two-stage divide-and-conquer training strategy termed TDC Training. It groups timesteps based on task similarity and difficulty, assigning highly customized denoising models to each group, thereby enhancing the performance of diffusion models. While two-stage training avoids the need to train each model separately, the total training cost is even lower than training a single unified denoising model. Additionally, we introduce Proxy-based Pruning to further customize the denoising models. This method transforms the pruning problem of diffusion models into a multi-round decision-making problem, enabling precise pruning of diffusion models. Our experiments validate the effectiveness of TDC Training, demonstrating improvements in FID of 1.5 on ImageNet64 compared to original IDDPM, while saving about 20% of computational resources.

In vision-language models (VLMs), visual tokens usually consume a significant amount of computational overhead, despite their sparser information density compared to text tokens. To address this, most existing methods learn a network to prune redundant visual tokens and require additional training data. Differently, we propose an efficient training-free token optimization mechanism dubbed SparseVLM without extra parameters or fine-tuning costs. Concretely, given that visual tokens complement text tokens in VLMs for linguistic reasoning, we select visual-relevant text tokens to rate the significance of vision tokens within the self-attention matrix extracted from the VLMs. Then we progressively prune irrelevant tokens. To maximize sparsity while retaining essential information, we introduce a rank-based strategy to adaptively determine the sparsification ratio for each layer, alongside a token recycling method that compresses pruned tokens into more compact representations. Experimental results show that our SparseVLM improves the efficiency of various VLMs across a range of image and video understanding tasks. In particular, when LLaVA is equipped with SparseVLM, it achieves a 54% reduction in FLOPs, lowers CUDA time by 37%, and maintains an accuracy rate of 97%. Our code is available at https://github.com/Gumpest/SparseVLMs.

Vision-Language-Action (VLA) models have demonstrated strong multi-modal reasoning capabilities, enabling direct action generation from visual perception and language instructions in an end-to-end manner. However, their substantial computational cost poses a challenge for real-time robotic control, where rapid decision-making is essential. This paper introduces VLA-Cache, a training-free inference acceleration method that reduces computational overhead by adaptively caching and reusing static visual tokens across frames. Exploiting the temporal continuity in robotic manipulation, VLA-Cache identifies minimally changed tokens between adjacent frames and reuses their cached key-value representations, thereby circumventing redundant computations. Additionally, to maintain action precision, VLA-Cache selectively re-computes task-relevant tokens that are environmentally sensitive, ensuring the fidelity of critical visual information. To further optimize efficiency, we introduce a layer adaptive token reusing strategy that dynamically adjusts the reuse ratio based on attention concentration across decoder layers, prioritizing critical tokens for recomputation. Extensive experiments on two simulation platforms (LIBERO and SIMPLER) and a real-world robotic system demonstrate that VLA-Cache achieves up to 1.7× speedup in CUDA latency and a 15\% increase in control frequency, with negligible loss on task success rate. The code and videos can be found at our project page: https://vla-cache.github.io.

The paper introduces DIST, an innovative knowledge distillation method that excels in learning from a superior teacher model. DIST differentiates itself from conventional techniques by adeptly handling the often significant prediction discrepancies between the student and teacher models. It achieves this by focusing on maintaining the relationships between their predictions, implementing a correlation-based loss to explicitly capture the teacher's intrinsic inter-class relations. Moreover, DIST uniquely considers the semantic similarities between different instances and each class at the intra-class level. The method is further enhanced by two significant improvements: (1) A teacher acclimation strategy, which effectively reduces the discrepancy between teacher and student, thereby optimizing the distillation process. (2) An extension of the DIST loss from the logit level to the feature level, a modification that proves especially beneficial for dense prediction tasks. DIST stands out for its simplicity, practicality, and adaptability to various architectures, model sizes, and training strategies. It consistently delivers state-of-the-art results across a range of applications, including image classification, object detection, and semantic segmentation.

Knowledge distillation (KD) has been applied to various tasks successfully, and mainstream methods typically boost the student model via spatial imitation losses. However, the consecutive downsamplings induced in the spatial domain of teacher model is a type of corruption, hindering the student from analyzing what specific information needs to be imitated, which results in accuracy degradation. To better understand the underlying pattern of corrupted feature maps, we shift our attention to the frequency domain. During frequency distillation, we encounter a new challenge: the low-frequency bands convey general but minimal context, while the high are more informative but also introduce noise. Not each pixel within the frequency bands contributes equally to the performance. To address the above problem: (1) We propose the Frequency Prompt plugged into the teacher model, absorbing the semantic frequency context during finetuning. (2) During the distillation period, a pixel-wise frequency mask is generated via Frequency Prompt, to localize those pixel of interests (PoIs) in various frequency bands. Additionally, we employ a position-aware relational frequency loss for dense prediction tasks, delivering a high-order spatial enhancement to the student model. We dub our Frequency Knowledge Distillation method as FreeKD, which determines the optimal localization and extent for the frequency distillation. Extensive experiments demonstrate that FreeKD not only outperforms spatial-based distillation methods consistently on dense prediction tasks (e.g., FreeKD brings 3.8 AP gains for RepPoints-R50 on COCO2017 and 4.55 mIoU gains for PSPNet-R18 on Cityscapes), but also conveys more robustness to the student. Notably, we also validate the generalization of our approach on large-scale vision models (e.g., DINO and SAM).

Large vision-language models (LVLMs) have recently achieved rapid progress, exhibiting great perception and reasoning abilities concerning visual information. However, when faced with prompts in different sizes of solution spaces, LVLMs fail to always give consistent answers regarding the same knowledge point. This inconsistency of answers between different solution spaces is prevalent in LVLMs and erodes trust. To this end, we provide a multi-modal benchmark ConBench, to intuitively analyze how LVLMs perform when the solution space of a prompt revolves around a knowledge point. Based on the ConBench tool, we are the first to reveal the tapestry and get the following findings: (1) In the discriminate realm, the larger the solution space of the prompt, the lower the accuracy of the answers. (2) Establish the relationship between the discriminative and generative realms: the accuracy of the discriminative question type exhibits a strong positive correlation with its Consistency with the caption. (3) Compared to open-source models, closed-source models exhibit a pronounced bias advantage in terms of Consistency. Eventually, we ameliorate the consistency of LVLMs by trigger-based diagnostic refinement, indirectly improving the performance of their caption. We hope this paper will accelerate the research community in better evaluating their models and encourage future advancements in the consistency domain.

Knowledge distillation is an effective paradigm for boosting the performance of pocket-size model, especially when multiple teacher models are available, the student would break the upper limit again. However, it is not economical to train diverse teacher models for the disposable distillation. In this paper, we introduce a new concept dubbed Avatars for distillation, which are the inference ensemble models derived from the teacher. Concretely, (1) For each iteration of distillation training, various Avatars are generated by a perturbation transformation. We validate that Avatars own higher upper limit of working capacity and teaching ability, aiding the student model in learning diverse and receptive knowledge perspectives from the teacher model. (2) During the distillation, we propose an uncertainty-aware factor from the variance of statistical differences between the vanilla teacher and Avatars, to adjust Avatars' contribution on knowledge transfer adaptively. Avatar Knowledge Distillation (AKD) is fundamentally different from existing methods and refines with the innovative view of unequal training. Comprehensive experiments demonstrate the effectiveness of our Avatars mechanism, which polishes up the state-of-the-art distillation methods for dense prediction without more extra computational cost. The AKD brings at most 0.7 AP gains on COCO 2017 for Object Detection and 1.83 mIoU gains on Cityscapes for Semantic Segmentation, respectively.

The representation gap between teacher and student is an emerging topic in knowledge distillation (KD). To reduce the gap and improve the performance, current methods often resort to complicated training schemes, loss functions, and feature alignments, which are task-specific and feature-specific. In this paper, we state that the essence of these methods is to discard the noisy information and distill the valuable information in the feature, and propose a novel KD method dubbed DiffKD, to explicitly denoise and match features using diffusion models. Our approach is based on the observation that student features typically contain more noises than teacher features due to the smaller capacity of student model. To address this, we propose to denoise student features using a diffusion model trained by teacher features. This allows us to perform better distillation between the refined clean feature and teacher feature. Additionally, we introduce a light-weight diffusion model with a linear autoencoder to reduce the computation cost and an adaptive noise matching module to improve the denoising performance. Extensive experiments demonstrate that DiffKD is effective across various types of features and achieves state-of-the-art performance consistently on image classification, object detection, and semantic segmentation tasks. Code is available at https://github.com/hunto/DiffKD.

Training a good supernet in one-shot NAS methods is difficult since the search space is usually considerably huge$(\mathrm{e}.\mathrm{g}.,\ 13^{21})$. In order to enhance the supernet's evaluation ability, one greedy strategy is to sample good paths, and let the supernet lean towards the good ones and ease its evaluation burden as a result. However, in practice the search can be still quite inefficient since the identification of good paths is not accurate enough and sampled paths still scatter around the whole search space. In this paper, we leverage an explicit path filter to capture the characteristics of paths and directly filter those weak ones, so that the search can be thus implemented on the shrunk space more greedily and efficiently. Concretely, based on the fact that good paths are much less than the weak ones in the space, we argue that the label of “weak paths” will be more confident and reliable than that of “good paths” in multi-path sampling. In this way, we thus cast the training of path filter in the positive and unlabeled (PU) learning paradigm, and also encourage a path embedding as better path/operation representation to enhance the identification capacity of the learned filter. By dint of this embedding, we can further shrink the search space by aggregating similar operations with similar embeddings, and the search can be more efficient and accurate. Extensive experiments validate the effectiveness of the proposed method GreedyNASv2. For example, our obtained GreedyNASv2-L achieves 81.1% Top-1 accuracy on ImageNet dataset, significantly outperforming the ResNet-50 strong baselines.

Structural re-parameterization (Rep) methods achieve noticeable improvements on simple VGG-style networks. Despite the prevalence, current Rep methods simply re-parameterize all operations into an augmented network, including those that rarely contribute to the model's performance. As such, the price to pay is an expensive computational overhead to manipulate these unnecessary behaviors. To eliminate the above caveats, we aim to boot-strap the training with minimal cost by devising a dynamic re-parameterization (DyRep) method, which encodes Rep technique into the training process that dynamically evolves the network structures. Concretely, our proposal adaptively finds the operations which contribute most to the loss in the network, and applies Rep to enhance their representational capacity. Besides, to suppress the noisy and redundant operations introduced by Rep, we devise a de-parameterization technique for a more compact re-parameterization. With this regard, DyRep is more efficient than Rep since it smoothly evolves the given network instead of constructing an over-parameterized network. Experimental results demonstrate our effectiveness, e.g., DyRep improves the accuracy of ResNet-18 by 2.04% on ImageNet and reduces 22% runtime over the baseline. Code is avail-able at: https://github.com/hunto/DyRep.

Filter pruning is essential for deploying a well-trained CNN model on edge computation devices with a target computation budget (e.g., FLOPs). Current filter pruning methods mainly focus on leveraging feature maps to analyze the importance of filters, and prune those with less impact on the value of the CNN’s loss function, thereby ignoring the variance of input batches to differences in sparse structure over the filters. In this paper, we propose a data-agnostic filter pruning method that uses an auxiliary network named Dagger module to induce pruning with the pre-trained weights as input. Besides, to help prune filters with a preset FLOPs constraint, we utilize an explicit FLOPs-aware regularisation mechanism to directly promote pruning filters toward the target FLOPs. Experimental results on CIFAR-10 and ImageNet datasets show that the proposed filter pruning method surpasses the state-of-the-art.

Unlike existing knowledge distillation methods focus on the baseline settings, where the teacher models and training strategies are not that strong and competing as state-of-the-art approaches, this paper presents a method dubbed DIST to distill better from a stronger teacher. We empirically find that the discrepancy of predictions between the student and a stronger teacher may tend to be fairly severer. As a result, the exact match of predictions in KL divergence would disturb the training and make existing methods perform poorly. In this paper, we show that simply preserving the relations between the predictions of teacher and student would suffice, and propose a correlation-based loss to capture the intrinsic inter-class relations from the teacher explicitly. Besides, considering that different instances have different semantic similarities to each class, we also extend this relational match to the intra-class level. Our method is simple yet practical, and extensive experiments demonstrate that it adapts well to various architectures, model sizes and training strategies, and can achieve state-of-the-art performance consistently on image classification, object detection, and semantic segmentation tasks. Code is available at: https://github.com/hunto/DIST_KD.

One-shot neural architecture search (NAS) methods significantly reduce the search cost by considering the whole search space as one network, which only needs to be trained once. However, current methods select each operation independently without considering previous layers. Besides, the historical information obtained with huge computation costs is usually used only once and then discarded. In this paper, we introduce a sampling strategy based on Monte Carlo tree search (MCTS) with the search space modeled as a Monte Carlo tree (MCT), which captures the dependency among layers. Furthermore, intermediate results are stored in the MCT for future decisions and a better exploration-exploitation balance. Concretely, MCT is updated using the training loss as a reward to the architecture performance; for accurately evaluating the numerous nodes, we propose node communication and hierarchical node selection methods in the training and search stages, respectively, making better uses of the operation rewards and hierarchical information. Moreover, for a fair comparison of different NAS methods, we construct an open-source NAS benchmark of a macro search space evaluated on CIFAR-10, namely NAS-Bench-Macro. Extensive experiments on NAS-Bench-Macro and ImageNet demonstrate that our method significantly improves search efficiency and performance. For example, by only searching 20 architectures, our obtained architecture achieves 78.0% top-1 accuracy with 442M FLOPs on ImageNet. Code (Benchmark) is available at: https://github.com/xiusu/NAS-Bench-Macro.

Training a supernet matters for one-shot neural architecture search (NAS) methods since it serves as a basic performance estimator for different architectures (paths). Current methods mainly hold the assumption that a supernet should give a reasonable ranking over all paths. They thus treat all paths equally, and spare much effort to train paths. However, it is harsh for a single supernet to evaluate accurately on such a huge-scale search space (e.g., 7^21). In this paper, instead of covering all paths, we ease the burden of supernet by encouraging it to focus more on evaluation of those potentially-good ones, which are identified using a surrogate portion of validation data. Concretely, during training, we propose a multi-path sampling strategy with rejection, and greedily filter the weak paths. The training efficiency is thus boosted since the training space has been greedily shrunk from all paths to those potentially-good ones. Moreover, we further adopt an exploration and exploitation policy by introducing an empirical candidate path pool. Our proposed method GreedyNAS is easy-to-follow, and experimental results on ImageNet dataset indicate that it can achieve better Top-1 accuracy under same search space and FLOPs or latency level, but with only ~60% of supernet training cost. By searching on a larger space, our GreedyNAS can also obtain new state-of-the-art architectures.