Speech-to-Speech (S2S) Large Language Models (LLMs) are foundational to natural human-computer interaction, enabling end-to-end spoken dialogue systems. However, evaluating these models remains a fundamental challenge. We propose SageLM, an end-to-end, multi-aspect, and explainable speech LLM for comprehensive S2S LLMs evaluation. First, unlike cascaded approaches that disregard acoustic features, SageLM jointly assesses both semantic and acoustic dimensions. Second, it leverages rationale-based supervision to enhance explainability and guide model learning, achieving superior alignment with evaluation outcomes compared to rule-based reinforcement learning methods. Third, we introduce SpeechFeedback, a synthetic preference dataset, and employ a two-stage training paradigm to mitigate the scarcity of speech preference data. Trained on both semantic and acoustic dimensions, SageLM achieves an 82.79% agreement rate with human evaluators, outperforming cascaded and SLM-based baselines by at least 7.42% and 26.20%, respectively.

Flow matching-based generative models offer a principled approach to modeling continuous-time dynamics in speech generation. However, inference is often computationally expensive due to repeated neural network evaluations required by ODE solvers. We propose WaveEx, a training-free and plug-in acceleration framework which replaces portions of ODE integration with wavelet-guided extrapolation. By leveraging the multi-scale structure of latent trajectories, WaveEx predicts future states directly in the frequency domain without additional model evaluations or architectural changes. WaveEx consistently accelerates inference across diverse speech generation tasks. The gains are especially pronounced in tasks like speech synthesis (up to 5.73× speedup) and music generation (2.75×), where flow matching plays a central role in alignment modeling and dense ODE integration. Even in tasks with simpler input-output mappings such as speech enhancement (4.55×) and voice conversion (2.75×), WaveEx still achieves notable acceleration, demonstrating the robustness and generalizability of the approach. These results highlight wavelet-guided extrapolation as a lightweight and broadly applicable alternative to full ODE solving for flow matching-based speech generation.

Data condensation techniques aim to synthesize a compact dataset from a larger one to enable efficient model training, yet while successful in unimodal settings, they often fail in multimodal scenarios where preserving intricate inter-modal dependencies is crucial. To address this, we introduce ImageBindDC, a novel data condensation framework operating within the unified feature space of ImageBind. Our approach moves beyond conventional distribution-matching by employing a powerful Characteristic Function (CF) loss, which operates in the Fourier domain to facilitate a more precise statistical alignment via exact infinite moment matching. We design our objective to enforce three critical levels of distributional consistency: (i) uni-modal alignment, which matches the statistical properties of synthetic and real data within each modality; (ii) cross-modal alignment, which preserves pairwise semantics by matching the distributions of hybrid real-synthetic data pairs; and (iii) joint-modal alignment, which captures the complete multivariate data structure by aligning the joint distribution of real data pairs with their synthetic counterparts. Extensive experiments highlight the effectiveness of ImageBindDC: on the NYU-v2 dataset, a model trained on just 5 condensed datapoints per class achieves lossless performance comparable to one trained on the full dataset, achieving a new state-of-the-art with an 8.2% absolute improvement over the previous best method and more than 4× less condensation time.

The growing scale of datasets in deep learning has introduced significant computational challenges. Dataset pruning addresses this challenge by constructing a compact but informative coreset from the full dataset with comparable performance. Previous approaches typically establish scoring metrics based on specific criteria to identify representative samples. However, these methods predominantly rely on sample scores obtained from the model's performance during the training (i.e., fitting) phase. As scoring models achieve near-optimal performance on training data, such fitting-centric approaches induce a dense distribution of sample scores within a narrow numerical range. This concentration reduces the distinction between samples and hinders effective selection. To address this challenge, we conduct dataset pruning from the perspective of generalization, i.e., scoring samples based on models not exposed to them during training. We propose a plug-and-play framework, UNSEEN, which can be integrated into existing dataset pruning methods. Additionally, conventional score-based methods are single-step and rely on models trained solely on the complete dataset, providing limited perspective on the importance of samples. To address this limitation, we scale UNSEEN to multi-step scenarios and propose an incremental selection technique through scoring models trained on varying coresets, and optimize the quality of the coreset dynamically. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art (SOTA) methods on CIFAR-10, CIFAR-100, and ImageNet-1K. Notably, on ImageNet-1K, UNSEEN achieves lossless performance while reducing training data by 30%.

Processing long visual token sequences poses a significant computational burden on Multimodal Large Language Models (MLLMs). While token pruning offers a path to acceleration, we find that current methods, while adequate for general understanding, catastrophically fail on fine-grained localization tasks. We attribute this failure to the inherent flaws of the two prevailing strategies: importance-based methods suffer from a strong positional bias, an inherent model artifact that distracts from semantic content, while diversity-based methods exhibit structural blindness, disregarding the user's prompt and spatial redundancy. To address this, we introduce D²Pruner, a framework that rectifies these issues by uniquely combining debiased importance with a structural pruning mechanism. Our method first secures a core set of the most critical tokens as pivots based on a debiased attention score. It then performs a Maximal Independent Set (MIS) selection on the remaining tokens, which are modeled on a hybrid graph where edges signify spatial proximity and semantic similarity. This process iteratively preserves the most important and available token while removing its neighbors, ensuring that the supplementary tokens are chosen to maximize importance and diversity. Extensive experiments demonstrate that D²Pruner achieves exceptional efficiency and fidelity.

Diffusion Transformers (DiTs) have demonstrated exceptional performance in high-fidelity image and video generation. To reduce their substantial computational costs, feature caching techniques have been proposed to accelerate inference by reusing hidden representations from previous timesteps. However, current methods often struggle to maintain generation quality at high acceleration ratios, where prediction errors increase sharply due to the inherent instability of long-step forecasting. In this work, we adopt an ordinary differential equation (ODE) perspective on the hidden-feature sequence, modeling layer representations along the trajectory as a feature-ODE. We attribute the degradation of existing caching strategies to their inability to robustly integrate historical features under large skipping intervals. To address this, we propose FoCa (Forecast-then-Calibrate), which treats feature caching as a feature-ODE solving problem. Extensive experiments on image, video generation, and super-resolution tasks demonstrate the effectiveness of FoCa, especially under aggressive acceleration. Without additional training, FoCa achieves near-lossless speedups of 5.50× on FLUX, 6.45× on HunyuanVideo, 3.17× on Inf-DiT, and maintains high quality with a 4.53× speedup on DiT.

Shuang Cheng, Yuhua Jiang, Zineng Zhou, Dawei Liu, Tao Wang, Linfeng Zhang, Biqing Qi, Bowen Zhou. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Chenfei Liao, Wensong Wang, Zichen Wen, Xu Zheng, Yiyu Wang, Haocong He, Yuanhuiyi Lyu, Lutao Jiang, Xin Zou, Yuqian Fu, Bin Ren, Linfeng Zhang, Xuming Hu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

With the advent of large vision-language models (LVLMs) demonstrating increasingly human-like abilities, a pivotal question emerges: do different LVLMs interpret multimodal sarcasm differently, and can a single model grasp sarcasm from multiple perspectives like humans? To explore this, we introduce an analytical framework using systematically designed prompts on existing multimodal sarcasm datasets. Evaluating 12 state-of-the-art LVLMs over 2409 samples, we examine interpretive variations within and across models, focusing on confidence levels, alignment with dataset labels, and recognition of ambiguous “neutral” cases. We further validate our findings on a diverse 100-sample mini-benchmark, incorporating multiple datasets, expanded prompt variants, and representative commercial LVLMs. Our findings reveal notable discrepancies—across LVLMs and within the same model under varied prompts. While classification-oriented prompts yield higher internal consistency, models diverge markedly when tasked with interpretive reasoning. These results challenge binary labeling paradigms by highlighting sarcasm’s subjectivity. We advocate moving beyond rigid annotation schemes toward multiperspective, uncertainty-aware modeling, offering deeper insights into multimodal sarcasm comprehension.

Diffusion Transformers (DiT) have become the dominant methods in image and video generation yet still suffer substantial computational costs. As an effective approach for DiT acceleration, feature caching methods are designed to cache the features of DiT in previous timesteps and reuse them in the next timesteps, allowing us to skip the computation in the next timesteps. Among them, token-wise feature caching has been introduced to perform different caching ratios for different tokens in DiTs, aiming to skip the computation for unimportant tokens while still computing the important ones. In this paper, we propose to carefully check the effectiveness in token-wise feature caching with the following two questions: 1) Is it really necessary to compute the so-called "important" tokens in each step? 2) Are so-called important tokens really important? Surprisingly, this paper gives some counter-intuition answers, demonstrating that consistently computing the selected "important tokens" in all steps is not necessary. The selection of the so-called "important tokens" is often ineffective, and even sometimes shows inferior performance than random selection. Based on these observations, this paper introduces dual feature caching referred to as DuCa, which performs aggressive caching strategy and conservative caching strategy iteratively and selects the tokens for computing randomly. Extensive experimental results demonstrate the effectiveness of our method in DiT, PixArt, FLUX, and OpenSora, demonstrating significant improvements than the previous token-wise feature caching.

Sarcasm in social media, frequently conveyed through the interplay of text and images, presents significant challenges for sentiment analysis and intention mining. Existing multi-modal sarcasm detection approaches have been shown to excessively depend on superficial cues within the textual modality, exhibiting limited capability to accurately discern sarcasm through subtle text–image interactions. To address this limitation, a novel framework, InterCLIP-MEP, is proposed. This framework integrates Interactive CLIP (InterCLIP), which employs an efficient training strategy to derive enriched cross-modal representations by embedding inter-modal information directly into each encoder, while using approximately 20.6 \(\times\) fewer trainable parameters compared with existing state-of-the-art (SOTA) methods. Furthermore, a Memory-Enhanced Predictor (MEP) is introduced, featuring a dynamic dual-channel memory mechanism that captures and retains valuable knowledge from test samples during inference, serving as a nonparametric classifier to enhance sarcasm detection robustness. Extensive experiments on MMSD, MMSD2.0, and DocMSU show that InterCLIP-MEP achieves SOTA performance, specifically improving accuracy by 1.08% and F1-score by 1.51% on MMSD2.0. Under distributional shift evaluation, it attains 73.96% accuracy, exceeding its memory-free variant by nearly 10% and the previous SOTA by over 15%, demonstrating superior stability and adaptability. The implementation of InterCLIP-MEP is publicly available at https://github.com/CoderChen01/InterCLIP-MEP .

The excellent performance of diffusion models in image generation is always accompanied by overlarge computation costs, which have prevented the application of diffusion models in edge devices and interactive applications. Previous works mainly focus on using fewer sampling steps and compressing the denoising network of diffusion models, while this paper proposes to accelerate diffusion models by introducing SiTo, a similarity-based token pruning method that adaptive prunes the redundant tokens in the input data. SiTo is designed to maximize the similarity between model prediction with and without token pruning by using cheap and hardware-friendly operations, leading to significant acceleration ratios without performance drop, and even sometimes improvements in the generation quality. For instance, the zero-shot evaluation shows SiTo leads to 1.90x and 1.75x acceleration on COCO30K and ImageNet with 1.33 and 1.15 FID reduction at the same time. Besides, SiTo has no training requirements and does not require any calibration data, making it plug-and-play in real-world applications.

Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model’s predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4× speedup.

Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: $\textbf{neuron misidentification}$ due to simplistic parameter magnitude-based selection, and $\textbf{cross-task neuron interference}$ during merging. To address these challenges, we propose $\textbf{LED-Merging}$, a three-stage framework that $\textbf{L}$ocates task-specific neurons via gradient-based attribution, dynamically $\textbf{E}$lects critical neurons through multi-model importance fusion, and $\textbf{D}$isjoints conflicting updates through parameter isolation. Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging effectively reduces harmful response rates, showing a 31.4\% decrease on Llama-3-8B-Instruct on HarmBench, while simultaneously preserving 95\% of utility performance, such as achieving 52.39\% accuracy on GSM8K. LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs. Code is available at $\href{https://github.com/MqLeet/LED-Merging}{GitHub}$.

Despite the recent breakthroughs achieved by Large Vision Language Models (LVLMs) in understanding and responding to complex visual-textual contexts, their inherent hallucination tendencies limit their practical application in real-world scenarios that demand high levels of precision. Existing methods typically either fine-tune the LVLMs using additional data, which incurs extra costs in manual annotation and computational resources or perform comparisons at the decoding stage, which may eliminate useful language priors for reasoning while introducing inference time overhead. Therefore, we propose ICT, a lightweight, training-free method that calculates an intervention direction to shift the model’s focus towards different levels of visual information, enhancing its attention to high-level and fine-grained visual details. During the forward pass stage, the intervention is applied to the attention heads that encode the overall image information and the fine-grained object details, effectively mitigating the phenomenon of overly language priors, and thereby alleviating hallucinations. Extensive experiments demonstrate that ICT achieves strong performance with a small amount of data and generalizes well across different datasets and models. Our codes are publicly available at:https://github.com/THU-BPM/ICT/.

Dataset distillation has emerged as a powerful approach for reducing data requirements in deep learning. Among various methods, distribution matching-based approaches stand out for their balance of computational efficiency and strong performance. However, existing distance metrics used in distribution matching often fail to accurately capture distributional differences, leading to unreliable measures of discrepancy. In this paper, we reformulate dataset distillation as a minmax optimization problem and introduce Neural Characteristic Function Discrepancy (NCFD), a comprehensive and theoretically grounded metric for measuring distributional differences. NCFD leverages the Characteristic Function (CF) to encapsulate full distributional information, employing a neural network to optimize the sampling strategy for the CF’s frequency arguments, thereby maximizing the discrepancy to enhance distance estimation. Simultaneously, we minimize the difference between real and synthetic data under this optimized NCFD measure. Our approach, termed Neural Characteristic Function Matching (NCFM), inherently aligns the phase and amplitude of neural features in the complex plane for both real and synthetic data, achieving a balance between realism and diversity in synthetic samples. Experiments demonstrate that our method achieves significant performance gains over state-of-the-art methods on both low- and high-resolution datasets. Notably, we achieve a 22.9% accuracy boost on ImageSquawk. Our method also reduces GPU memory usage by over 300× and achieves 20× faster processing speeds compared to state-of-the-art methods. To the best of our knowledge, this is the first work to achieve lossless compression of CIFAR-100 on a single NVIDIA 2080 Ti GPU using only 2.3 GB of memory. The code for this work is publicly available at: https://github.com/gszfwsb/NCFM.

Diffusion models have achieved significant progress in both image and video generation while still suffering from huge computation costs. As an effective solution, rectified flow aims to rectify the diffusion process of diffusion models into a straight line for few-step and even one-step generation. However, in this paper, we suggest that the original training pipeline of reflow is not optimal and introduce two techniques to improve it. Firstly, we introduce progressive reflow, which progressively reflows the diffusion models in local timesteps until the whole diffusion progresses, reducing the difficulty of flow matching. Second, we introduce aligned v-prediction, which highlights the importance of direction matching in flow matching over magnitude matching. Experimental results on SDv1.5 and SDXL demonstrate the effectiveness of our method, for example, conducting on SDv1.5 achieves an FID of 10.70 on MSCOCO2014 validation set with only 4 sampling steps, close to our teacher model (32 DDIM steps, FID = 10.05). Our codes will be released at Github.

Diffusion models are trained by learning a sequence of models that reverse each step of noise corruption. Typically, the model parameters are fully shared across multiple timesteps to enhance training efficiency. However, since the denoising tasks differ at each timestep, the gradients computed at different timesteps may conflict, potentially degrading the overall performance of image generation. To solve this issue, this work proposes a Decouple-then-Merge (DeMe) framework, which begins with a pretrained model and finetunes separate models tailored to specific timesteps. We introduce several improved techniques during the fine-tuning stage to promote effective knowledge sharing while minimizing training interference across timesteps. Finally, after finetuning, these separate models can be merged into a single model in the parameter space, ensuring efficient and practical inference. Experimental results show significant generation quality improvements upon 6 benchmarks including Stable Diffusion on COCO30K, ImageNet1K, PartiPrompts, and DDPM on LSUN Church, LSUN Bedroom, and CIFAR10. Code is available at GitHub.

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance.Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference.To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression.In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships.Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens.GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-andplay manner.Our code is publicly available at GitHub.

Vision tokens in multimodal large language models often dominate huge computational overhead due to their excessive length compared to linguistic modality. Abundant recent methods aim to solve this problem with token pruning, which first defines an importance criterion for tokens and then prunes the unimportant vision tokens during inference. However, in this paper, we show that the importance is not an ideal indicator to decide whether a token should be pruned. Surprisingly, it usually results in inferior performance than random token pruning and leading to incompatibility to efficient attention computation operators. Instead, we propose DART (Duplication-Aware Reduction of Tokens), which prunes tokens based on its duplication with other tokens, leading to significant and training-free acceleration. Concretely, DART selects a small subset of pivot tokens and then retains the tokens with low duplication to the pivots, ensuring minimal information loss during token pruning. Experiments demonstrate that DART can prune 88.9% vision tokens while maintaining comparable performance, leading to a 1.99\times and 2.99\times speed-up in total time and prefilling stage, respectively, with good compatibility to efficient attention operators.

The rapid advancements in generative technology have emerged as a double-edged sword. While offering powerful tools that enhance convenience, they also pose significant social concerns. As defenders, current synthetic image detection methods often lack artifact-level textual interpretability and are overly focused on image manipulation detection, and current datasets usually suffer from outdated generators and a lack of fine-grained annotations. In this paper, we introduce SynthScars, a high-quality and diverse dataset consisting of 12,236 fully synthetic images with human-expert annotations. It features 4 distinct image content types, 3 categories of artifacts, and fine-grained annotations covering pixel-level segmentation, detailed textual explanations, and artifact category labels. Furthermore, we propose LEGION (LEarning to Ground and explain for Synthetic Image detectiON), a multimodal large language model (MLLM)-based image forgery analysis framework that integrates artifact detection, segmentation, and explanation. Building upon this capability, we further explore LEGION as a controller, integrating it into image refinement pipelines to guide the generation of higher-quality and more realistic images. Extensive experiments show that LEGION outperforms existing methods across multiple benchmarks, particularly surpassing the second-best traditional expert on SynthScars by 3.31% in mIoU and 7.75% in F1 score. Moreover, the refined images generated under its guidance exhibit stronger alignment with human preferences. The code, model, and dataset will be released.

Diffusion Transformers (DiT) have revolutionized high-fidelity image and video synthesis, yet their computational demands remain prohibitive for real-time applications. To solve this problem, feature caching has been proposed to accelerate diffusion models by caching the features in the previous timesteps and then reusing them in the following timesteps. However, at timesteps with significant intervals, the feature similarity in diffusion models decreases substantially, leading to a pronounced increase in errors introduced by feature caching, significantly harming the generation quality. To solve this problem, we propose TaylorSeer, which firstly shows that features of diffusion models at future timesteps can be predicted based on their values at previous timesteps. Based on the fact that features change slowly and continuously across timesteps, TaylorSeer employs a differential method to approximate the higher-order derivatives of features and predict features in future timesteps with Taylor series expansion. Extensive experiments demonstrate its significant effectiveness in both image and video synthesis, especially in high acceleration ratios. For instance, it achieves an almost lossless acceleration of 4.99$\times$ on FLUX and 5.00$\times$ on HunyuanVideo without additional training. On DiT, it achieves $3.41$ lower FID compared with previous SOTA at $4.53$$\times$ acceleration. %Our code is provided in the supplementary materials and will be made publicly available on GitHub. Our codes have been released in Github:https://github.com/Shenyi-Z/TaylorSeer

The debate between self-interpretable models and post-hoc explanations for black-box models is central to Explainable AI (XAI). Self-interpretable models, such as concept-based networks, offer insights by connecting decisions to human-understandable concepts but often struggle with performance and scalability. Conversely, post-hoc methods like Shapley values, while theoretically robust, are computationally expensive and resource-intensive. To bridge the gap between these two lines of research, we propose a novel method that combines their strengths, providing theoretically guaranteed self-interpretability for black-box models without compromising prediction accuracy. Specifically, we introduce a parameter-efficient pipeline, AutoGnothi, which integrates a small side network into the black-box model, allowing it to generate Shapley value explanations without changing the original network parameters. This side-tuning approach significantly reduces memory, training, and inference costs, outperforming traditional parameter-efficient methods, where full fine-tuning serves as the optimal baseline. AutoGnothi enables the black-box model to predict and explain its predictions with minimal overhead. Extensive experiments show that AutoGnothi offers accurate explanations for both vision and language tasks, delivering superior computational efficiency with comparable interpretability.

Protecting the intellectual property of open-source Large Language Models (LLMs) is very important, because training LLMs costs extensive computational resources and data. Therefore, model owners and third parties need to identify whether a suspect model is a subsequent development of the victim model. To this end, we propose a training-free REEF to identify the relationship between the suspect and victim models from the perspective of LLMs' feature representations. Specifically, REEF computes and compares the centered kernel alignment similarity between the representations of a suspect model and a victim model on the same samples. This training-free REEF does not impair the model's general capabilities and is robust to sequential fine-tuning, pruning, model merging, and permutations. In this way, REEF provides a simple and effective way for third parties and models' owners to protect LLMs' intellectual property together. Our code is publicly accessible at https://github.com/AI45Lab/REEF.

Diffusion transformers have shown significant effectiveness in both image and video synthesis at the expense of huge computation costs. To address this problem, feature caching methods have been introduced to accelerate diffusion transformers by caching the features in previous timesteps and reusing them in the following timesteps. However, previous caching methods ignore that different tokens exhibit different sensitivities to feature caching, and feature caching on some tokens may lead to 10$\times$ more destruction to the overall generation quality compared with other tokens. In this paper, we introduce token-wise feature caching, allowing us to adaptively select the most suitable tokens for caching, and further enable us to apply different caching ratios to neural layers in different types and depths. Extensive experiments on PixArt-alpha, OpenSora, and DiT demonstrate our effectiveness in both image and video generation with no requirements for training. For instance, 2.36$\times$ and 1.93$\times$ acceleration are achieved on OpenSora and PixArt-$\alpha$ with almost no drop in generation quality. Codes have been released in the supplementary material and Github.

Recent text-to-image generative models, e.g., Stable Diffusion V3 and Flux, have achieved notable progress. However, these models are strongly restricted to their limited knowledge, a.k.a., their own fixed parameters, that are trained with closed datasets. This leads to significant hallucinations or distortions when facing fine-grained and unseen novel real-world objects, e.g., the appearance of the Tesla Cybertruck. To this end, we present the first real-object-based retrieval-augmented generation framework (RealRAG), which augments fine-grained and unseen novel object generation by learning and retrieving real-world images to overcome the knowledge gaps of generative models. Specifically, to integrate missing memory for unseen novel object generation, we train a reflective retriever by self-reflective contrastive learning, which injects the generator’s knowledge into the sef-reflective negatives, ensuring that the retrieved augmented images compensate for the model’s missing knowledge. Furthermore, the real-object-based framework integrates fine-grained visual knowledge for the generative models, tackling the distortion problem and improving the realism for fine-grained object generation. Our Real-RAG is superior in its modular application to all types of state-of-the-art text-to-image generative models and also delivers remarkable performance boosts with all of them, such as a gain of 16.18% FID score with the auto-regressive model on the Stanford Car benchmark.

Diffusion models have revolutionized high-fidelity image and video synthesis, yet their computational demands remain prohibitive for real-time applications. These models face two fundamental challenges: strict temporal dependencies preventing parallelization, and computationally intensive forward passes required at each denoising step. Drawing inspiration from speculative decoding in large language models, we present SpeCa, a novel 'Forecast-then-verify' acceleration framework that effectively addresses both limitations. SpeCa's core innovation lies in introducing Speculative Sampling to diffusion models, predicting intermediate features for subsequent timesteps based on fully computed reference timesteps. Our approach implements a parameter-free verification mechanism that efficiently evaluates prediction reliability, enabling real-time decisions to accept or reject each prediction while incurring negligible computational overhead. Furthermore, SpeCa introduces sample-adaptive computation allocation that dynamically modulates resources based on generation complexity, allocating reduced computation for simpler samples while preserving intensive processing for complex instances. Experiments demonstrate 6.34x acceleration on FLUX with minimal quality degradation (5.5% drop), 7.3x speedup on DiT while preserving generation fidelity, and 79.84% VBench score at 6.1x acceleration for HunyuanVideo. The verification mechanism incurs minimal overhead (1.67%-3.5% of full inference costs), establishing a new paradigm for efficient diffusion model inference while maintaining generation quality even at aggressive acceleration ratios. Our codes have been released in Github: \textbf{https://github.com/Shenyi-Z/Cache4Diffusion}

Diffusion transformers have gained significant attention in recent years for their ability to generate high-quality images and videos, yet still suffer from a huge computational cost due to their iterative denoising process. Recently, feature caching has been introduced to accelerate diffusion transformers by caching the feature computation in previous timesteps and reusing it in the following timesteps, which leverage the temporal similarity of diffusion models while ignoring the similarity in the spatial dimension. In this paper, we introduce Cluster-Driven Feature Caching (ClusCa) as an orthogonal and complementary perspective for previous feature caching. Specifically, ClusCa performs spatial clustering on tokens in each timestep, computes only one token in each cluster and propagates their information to all the other tokens, which is able to reduce the number of tokens by over 90%. Extensive experiments on DiT, FLUX and HunyuanVideo demonstrate its effectiveness in both text-to-image and text-to-video generation. Besides, it can be directly applied to any diffusion transformer without requirements for training. For instance, ClusCa achieves 4.96x acceleration on FLUX with an ImageReward of 99.49%, surpassing the original model by 0.51%. The code is available at https://github.com/zhixin-zheng/ClusCa.

Visual tokens consume substantial computational resources in multi-modal large models (MLLMs), significantly compromising their efficiency. Recent works have attempted to improve efficiency by compressing visual tokens during training, either through modifications to model components or by introducing additional parameters. However, they often overlook the increased learning difficulty caused by such compression, as the model’s parameter space struggles to quickly adapt to the substantial perturbations in the feature space induced by token compression. In this work, we propose to develop Efficient MLLMs via Progressive Consistency Distillation (EPIC), a progressive learning framework. Specifically, by decomposing the feature space perturbations introduced by token compression along the token-wise and layer-wise dimensions, we introduce token consistency distillation and layer consistency distillation, respectively, aiming to reduce the training difficulty by leveraging guidance from a teacher model and following a progressive learning trajectory. Extensive experiments demonstrate the superior effectiveness, robustness, and generalization capabilities of our proposed framework.

Vision-Language-Action (VLA) models, particularly diffusion-based architectures, demonstrate transformative potential for embodied intelligence but are severely hampered by high computational and memory demands stemming from extensive inherent and inference-time redundancies. While existing acceleration efforts often target isolated inefficiencies, such piecemeal solutions typically fail to holistically address the varied computational and memory bottlenecks across the entire VLA pipeline, thereby limiting practical deployability. We introduce EfficientVLA, a structured and training-free inference acceleration framework that systematically eliminates these barriers by cohesively exploiting multifaceted redundancies. EfficientVLA synergistically integrates three targeted strategies: (1) pruning of functionally inconsequential layers from the language module, guided by an analysis of inter-layer redundancies; (2) optimizing the visual processing pathway through a task-aware strategy that selects a compact, diverse set of visual tokens, balancing task-criticality with informational coverage; and (3) alleviating temporal computational redundancy within the iterative diffusion-based action head by strategically caching and reusing key intermediate features. We apply our method to a standard VLA model CogACT, yielding a $1.93\times$ inference speedup and reduces FLOPs to 28.9%, with only a 0.6% success rate drop in the SIMPLER benchmark.

Despite their powerful capabilities, multimodal large language models (MLLMs) suffer from considerable computational overhead due to their reliance on massive visual tokens. Recent studies have explored token pruning to alleviate this problem, which typically uses text-vision cross-attention or [CLS] attention to assess and discard redundant visual tokens. In this work, we identify a critical limitation of such attention-first pruning approaches, i.e., they tend to preserve semantically similar tokens, resulting in pronounced performance drops under high pruning rates. To this end, we propose HoloV, a simple yet effective, plug-and-play visual token pruning framework for efficient inference. Distinct from previous attention-first schemes, HoloV rethinks token retention from a holistic perspective. By adaptively distributing the pruning budget across different spatial crops, HoloV ensures that the retained tokens capture the global visual context rather than isolated salient features. This strategy minimizes representational collapse and maintains task-relevant information even under aggressive pruning. Experimental results demonstrate that our HoloV achieves superior performance across various tasks, MLLM architectures, and pruning ratios compared to SOTA methods. For instance, LLaVA1.5 equipped with HoloV preserves 95.8% of the original performance after pruning 88.9% of visual tokens, achieving superior efficiency-accuracy trade-offs.

Magnetic microrobots demonstrate significant potential in medical applications by providing innovative solutions for precise treatment through targeted drug delivery, minimally invasive surgery, and vascular cleaning. However, within biological organisms, there are various complex obstacle environments that require a navigation technology prioritizing safety and emphasizing smoothness. This paper proposes a safety-enhanced navigation planning (SENP) algorithm to achieve multiple objectives such as safety, path smoothness, and short distance, enabling collision-free navigation in complex medical environments. Unlike traditional methods that require multiple heuristic cost functions to guide the navigation planning algorithm, our approach leverages the safety-enhancing features of the safe artificial potential field (SAPF) to promote collision-free navigation in complex medical environments. By adopting a goal-biased strategy to guide the generation of random sampling points, the number of iterations is reduced, and the convergence speed of the algorithm is improved. In addition, the traditional artificial potential field planning method often leads to the oscillation of the corridor, our method effectively addresses this issue along with the uncertainty in the quality of the initial path and the lengthy convergence time to the optimal path. Comparative analysis with various algorithms in different environments shows that our proposed method excels in terms of smoothness and path length under the premise of safety, making it suitable for magnetic microrobots in complex environments.Note to Practitioners—The motivation for this work lies in advancing safe, efficient, and rapid navigation strategies for magnetic microrobots in medical applications. While significant progress has been made in the development of magnetic microrobots, navigating through complex environments such as human blood vessels remains a substantial challenge. The ability to perform safe and effective motion planning within narrow and intricate channels is crucial for medical applications. In response to this need, we propose a path planning method specifically designed for magnetically actuated microrobots, based on SENP. Our approach emphasizes achieving a path that is not only short and smooth but also prioritizes safety throughout the navigation process. Compared to traditional sampling-based algorithms, our method effectively overcomes limitations such as initial path quality uncertainty and prolonged convergence to optimal paths. This allows for the rapid generation of a high-quality initial path with a focus on safety while significantly accelerating convergence. Moreover, our approach provides a robust solution for navigating narrow channels, making it highly suitable for challenging medical environments.

Diffusion Transformers (DiT) excel at image and video generation but face computational challenges due to the quadratic complexity of self-attention operators. We propose DiTFastAttn, a post-training compression method to alleviate the computational bottleneck of DiT. We identify three key redundancies in the attention computation during DiT inference: (1) spatial redundancy, where many attention heads focus on local information; (2) temporal redundancy, with high similarity between the attention outputs of neighboring steps; (3) conditional redundancy, where conditional and unconditional inferences exhibit significant similarity. We propose three techniques to reduce these redundancies: (1) $\textit{Window Attention with Residual Sharing}$ to reduce spatial redundancy; (2) $\textit{Attention Sharing across Timesteps}$ to exploit the similarity between steps; (3) $\textit{Attention Sharing across CFG}$ to skip redundant computations during conditional generation.

In-air gesture control extends a touch screen and enables contactless interaction, thus has become a popular research direction in the past few years. Prior work has implemented this functionality based on cameras, acoustic signals, and Wi-Fi via existing hardware on commercial devices. However, these methods have low user acceptance. Solutions based on cameras and acoustic signals raise privacy concerns, while WiFi-based solutions are vulnerable to background noise. As a result, these methods are not commercialized and recent flagship smartphones have implemented in-air gesture recognition by adding extra hardware on-board, such as mmWave radar and depth camera. The question is, can we support in-air gesture control on legacy devices without any hardware modifications? To answer this question, in this work, we propose, an in-air gesture recognition system leveraging the screen and ambient light sensor (ALS), which are ordinary modalities on mobile devices. For the transmitter side, we design a screen display mechanism to embed spatial information and preserve the viewing experience; for the receiver side, we develop a framework to recognize gestures from low-quality ALS readings. We implement and evaluate on both a tablet and several smartphones. Results show that can recognize$9$types of frequently used in-air gestures with an average accuracy of$96.1\%$.

The remarkable breakthroughs in point cloud representation learning have boosted their usage in real-world applications such as self-driving cars and virtual reality. However, these applications usually have a strict requirement for not only accurate but also efficient 3D object detection. Recently, knowledge distillation has been proposed as an effective model compression technique, which transfers the knowledge from an over-parameterized teacher to a lightweight student and achieves consistent effectiveness in 2D vision. However, due to point clouds' sparsity and irregularity, directly applying previous image-based knowledge distillation methods to point cloud detectors usually leads to unsatisfactory performance. To fill the gap, this paper proposes PointDistiller; a structured knowledge distillation framework for point clouds-based 3D detection. Concretely, PointDistiller includes local distillation which extracts and distills the local geometric structure of point clouds with dynamic graph convolution and reweighted learning strategy, which highlights student learning on the crucial points or voxels to improve knowledge distillation efficiency. Extensive experiments on both voxels-based and raw points-based detectors have demonstrated the effectiveness of our method over seven previous knowledge distillation methods. For instance, our 4 x compressed PointPillars student achieves 2.8 and 3.4 mAP improvements on BEV and 3D object detection, outperforming its teacher by 0.9 and 1.8 mAP, respectively. Codes are available in https://github.com/RunpeiDong/PointDistiller.

Knowledge distillation conducts an effective model compression method while holding some limitations: (1) the feature based distillation methods only focus on distilling the feature map but are lack of transferring the relation of data examples; (2) the relational distillation methods are either limited to the handcrafted functions for relation extraction, such as L2norm, or weak in inter- and intra- class relation modeling. Besides, the feature divergence of heterogeneous teacher-student architectures may lead to inaccurate relational knowledge transferring. In this work, we propose a novel training framework named Class-Oriented Relational Self Distillation (CORSD) to address the limitations. The trainable relation networks are designed to extract relation of structured data input, and they enable the whole model to better classify samples by transferring the relational knowledge from the deepest layer of the model to shallow layers. Besides, auxiliary classifiers are proposed to make relation networks capture class-oriented relation that benefits classification task. Experiments demonstrate that CORSD achieves remarkable improvements. Compared to baseline, 3.8%, 1.5% and 4.5% averaged accuracy boost can be observed on CIFAR100, ImageNet and CUB-200-2011, respectively.

CNN’s limited receptive field restricts its ability to capture long-range spatial-temporal dependencies, leading to unsatisfactory performance in video super-resolution (VSR). To tackle this challenge, this paper presents a novel multi-frequency representation enhancement module (MFE) that performs spatial-temporal information aggregation in the frequency domain. Specifically, MFE mainly includes a spatial-frequency representation enhancement branch which captures the long-range dependency in the spatial dimension, and an energy frequency representation enhancement branch to obtain the inter-channel feature relationship. Moreover, a novel model training method named privilege training is proposed to encode the privilege information from high-resolution videos to facilitate model training. With these two methods, we introduce a new VSR model named MFPI, which outperforms state-of-the-art methods by a large margin while maintaining good efficiency on various datasets, including REDS4, Vimeo, Vid4, and UDM10.

Significant advancements have been accomplished with deep neural networks in diverse visual tasks, which have substantially elevated their deployment in edge device software. However, during the update of neural network-based software, users are required to download all the parameters of the neural network anew, which harms the user experience. Motivated by previous progress in model compression, we propose a novel training methodology named Tiny Updater to address this issue. Specifically, by adopting the variant of pruning and knowledge distillation methods, Tiny Updater can update the neural network-based software by only downloading a few parameters (10%~20%) instead of all the parameters in the neural network. Experiments on eleven datasets of three tasks, including image classification, image-to-image translation, and video recognition have demonstrated its effectiveness. Codes have been released in https://github.com/ArchipLab-LinfengZhang/TinyUpdater.

Voxel-based methods have achieved state-of-the-art performance for 3D object detection in autonomous driving. However, their significant computational and memory costs pose a challenge for their application to resource-constrained vehicles. One reason for this high resource consumption is the presence of a large number of redundant background points in Lidar point clouds, resulting in spatial redundancy in both 3D voxel and BEV map representations. To address this issue, we propose an adaptive inference framework called Ada3D, which focuses on reducing the spatial redundancy to compress the model’s computational and memory cost. Ada3D adaptively filters the redundant input, guided by a lightweight importance predictor and the unique properties of the Lidar point cloud. Additionally, we maintain the BEV features’ intrinsic sparsity by introducing the Sparsity Preserving Batch Normalization. With Ada3D, we achieve 40% reduction for 3D voxels and decrease the density of 2D BEV feature maps from 100% to 20% without sacrificing accuracy. Ada3D reduces the model computational and memory cost by 5×, and achieves 1.52× / 1.45× end-to-end GPU latency and 1.5× / 4.5× GPU peak memory optimization for the 3D and 2D backbone respectively.

The excellent performance of deep neural networks is usually accompanied by a large number of parameters and computations, which have limited their usage on the resource- limited edge devices. To address this issue, abundant methods such as pruning, quantization and knowledge distillation have been proposed to compress neural networks and achieved significant breakthroughs. However, most of these compression methods focus on the architecture or the training method of neural networks but ignore the influence from data augmentation. In this paper, we revisit the usage of data augmentation in model compression and give a comprehensive study on the relation between model sizes and their optimal data augmentation policy. To sum up, we mainly have the following three observations: (A) Models in different sizes prefer data augmentation with different magnitudes. Hence, in iterative pruning, data augmentation with varying magnitudes leads to better performance than data augmentation with a consistent magnitude. (B) Data augmentation with a high magnitude may significantly improve the performance of large models but harm the performance of small models. Fortunately, small models can still benefit from strong data augmentations by firstly learning them with “additional parameters” and then discard these “additional parameters” during inference. (C) The prediction of a pre-trained large model can be utilized to measure the difficulty of data augmentation. Thus it can be utilized as a criterion to design better data augmentation policies. We hope this paper may promote more research on the usage of data augmentation in model compression.

Knowledge distillation, which aims to transfer the knowledge learned by a cumbersome teacher model to a lightweight student model, has become one of the most popular and effective techniques in computer vision. However, many previous knowledge distillation methods are designed for image classification and fail in more challenging tasks such as object detection. In this paper, we first suggest that the failure of knowledge distillation on object detection is mainly caused by two reasons: (1) the imbalance between pixels of foreground and background and (2) lack of knowledge distillation on the relation among different pixels. Then, we propose a structured knowledge distillation scheme, includingattention-guided distillationandnon-local distillationto address the two issues, respectively. Attention-guided distillation is proposed to find the crucial pixels of foreground objects with an attention mechanism and then make the students take more effort to learn their features. Non-local distillation is proposed to enable students to learn not only the feature of an individual pixel but also the relation between different pixels captured by non-local modules. Experimental results have demonstrated the effectiveness of our method on thirteen kinds of object detection models with twelve comparison methods for both object detection and instance segmentation. For instance, Faster RCNN with our distillation achieves 43.9 mAP on MS COCO2017, which is 4.1 higher than the baseline. Additionally, we show that our method is also beneficial to the robustness and domain generalization ability of detectors. Codes and model weights have been released on GitHub†

Data augmentation, which improves the diversity of datasets by applying image transformations, has become one of the most effective techniques in visual representation learning. Usually, the design of augmentation policies faces a diversity-difficulty trade-off. On the one hand, a simple augmentation leads to a low training set diversity, which can not improve model performance significantly. On the other hand, an excessively hard augmentation has an overlarge regularization effect which harms model performance. Recently, automatic augmentation methods have been proposed to address this issue by searching the optimal data augmentation policy from a predefined searching space. However, these methods still suffer from heavy searching overhead or complex optimization objectives. In this paper, instead of searching the optimal augmentation policy, we propose to break the diversity-difficulty trade-off from a multi-task learning perspective. By formulating model learning on the augmented images and the original images as the auxiliary task and the primary task in multi-task learning respectively, the hard augmentation does not directly influence the training of the primary branch and thus its negative influence can be alleviated. Hence, neural networks can learn valuable semantic information even with a totally random augmentation policy. Experimental results on ten datasets for four tasks demonstrate the superiority of our method over the other twelve methods. Codes have been released inhttps://github.com/ArchipLab-LinfengZhang/data-augmentation-multi-task.

Remarkable achievements have been attained with Generative Adversarial Networks (GANs) in image-to-image translation. However, due to a tremendous amount of parameters, state-of-the-art GANs usually suffer from low efficiency and bulky memory usage. To tackle this challenge, firstly, this paper investigates GANs performance from a frequency perspective. The results show that GANs, especially small GANs lack the ability to generate high-quality high frequency information. To address this problem, we propose a novel knowledge distillation method referred to as wavelet knowledge distillation. Instead of directly distilling the generated images of teachers, wavelet knowledge distillation first decomposes the images into different frequency bands with discrete wavelet transformation and then only distills the high frequency bands. As a result, the student GAN can pay more attention to its learning on high frequency bands. Experiments demonstrate that our method leads to 7.08× compression and 6.80× acceleration on CycleGAN with almost no performance drop. Additionally, we have studied the relation between discriminators and generators which shows that the compression of discriminators can promote the performance of compressed generators.

Quantized neural networks typically require smaller memory footprints and lower computation complexity, which is crucial for efficient deployment. However, quantization inevitably leads to a distribution divergence from the original network, which generally degrades the performance. To tackle this issue, massive efforts have been made, but most existing approaches lack statistical considerations and depend on several manual configurations. In this paper, we present an adaptive-mapping quantization method to learn an optimal latent sub-distribution that is inherent within models and smoothly approximated with a concrete Gaussian Mixture (GM). In particular, the network weights are projected in compliance with the GM-approximated sub-distribution. This sub-distribution evolves along with the weight update in a co-tuning schema guided by the direct task-objective optimization. Sufficient experiments on image classification and object detection over various modern architectures demonstrate the effectiveness, generalization property, and transferability of the proposed method. Besides, an efficient deployment flow for the mobile CPU is developed, achieving up to 7.46$\times$ inference acceleration on an octa-core ARM CPU. Our codes have been publicly released at https://github.com/RunpeiDong/DGMS.

Remarkable achievements have been obtained by deep neural networks in the last several years. However, the breakthrough in neural networks accuracy is always accompanied by explosive growth of computation and parameters, which leads to a severe limitation of model deployment. In this paper, we propose a novel knowledge distillation technique named self-distillation to address this problem. Self-distillation attaches several attention modules and shallow classifiers at different depths of neural networks and distills knowledge from the deepest classifier to the shallower classifiers. Different from the conventional knowledge distillation methods where the knowledge of the teacher model is transferred to another student model, self-distillation can be considered as knowledge transfer in the same model - from the deeper layers to the shallow layers. Moreover, the additional classifiers in self-distillation allow the neural network to work in a dynamic manner, which leads to a much higher acceleration. Experiments demonstrate that self-distillation has consistent and significant effectiveness on various neural networks and datasets. On average, 3.49 and 2.32 percent accuracy boost are observed on CIFAR100 and ImageNet. Besides, experiments show that self-distillation can be combined with other model compression methods, including knowledge distillation, pruning and lightweight model design.

Large deep neural network (DNN) models pose the key challenge to energy efficiency due to the significantly higher energy consumption of off-chip DRAM accesses than arithmetic or SRAM operations. It motivates the intensive research on model compression with two main approaches. Weight pruning leverages the redundancy in the number of weights and can be performed in a non-structured, which has higher flexibility and pruning rate but incurs index accesses due to irregular weights, or structured manner, which preserves the full matrix structure with a lower pruning rate. Weight quantization leverages the redundancy in the number of bits in weights. Compared to pruning, quantization is much more hardware-friendly and has become a "must-do" step for FPGA and ASIC implementations. Thus, any evaluation of the effectiveness of pruning should be on top of quantization. The key open question is, with quantization, what kind of pruning (non-structured versus structured) is most beneficial? This question is fundamental because the answer will determine the design aspects that we should really focus on to avoid the diminishing return of certain optimizations. This article provides a definitive answer to the question for the first time. First, we build ADMM-NN-S by extending and enhancing ADMM-NN, a recently proposed joint weight pruning and quantization framework, with the algorithmic supports for structured pruning, dynamic ADMM regulation, and masked mapping and retraining. Second, we develop a methodology for fair and fundamental comparison of non-structured and structured pruning in terms of both storage and computation efficiency. Our results show that ADMM-NN-S consistently outperforms the prior art: 1) it achieves 348× , 36× , and 8× overall weight pruning on LeNet-5, AlexNet, and ResNet-50, respectively, with (almost) zero accuracy loss and 2) we demonstrate the first fully binarized (for all layers) DNNs can be lossless in accuracy in many cases. These results provide a strong baseline and credibility of our study. Based on the proposed comparison framework, with the same accuracy and quantization, the results show that non-structured pruning is not competitive in terms of both storage and computation efficiency. Thus, we conclude that structured pruning has a greater potential compared to non-structured pruning. We encourage the community to focus on studying the DNN inference acceleration with structured sparsity.

It is well acknowledged in image processing domain that the information can be decomposed into different frequency parts and each part has its own merits. However, existing neural networks always ignore the distinctions and straightforwardly feed all the information into neural networks together, treating them equally. In this paper, we propose a novel neural networks framework named J-Net that decomposes images into different frequency bands and then processes them sequentially. Concretely, the images have been decomposed by wavelet transformation and then the wavelet coefficients are fed into neural networks gradually in different depth according to their decomposition levels. An attention module is utilized to facilitate the fusion of neural network features and injected information, yielding significant performance gain. Furthermore, we show how does the information with different frequency impact the accuracy of neural networks. Experiments show that 5.91%, 5.32% and 2.00% accuracy improvements on Caltech 101, Caltech256 and ImageNet, respectively.

In-air gesture control extends a touch screen and enables contactless interaction, thus has become a popular research direction in the past few years. Prior work has implemented this functionality based on cameras, acoustic signals, and Wi-Fi via existing hardware on commercial devices. However, these methods have low user acceptance. Solutions based on cameras and acoustic signals raise privacy concerns, while WiFi-based solutions are vulnerable to background noise. As a result, these methods are not commercialized and recent flagship smartphones have implemented in-air gesture recognition by adding extra hardware on-board, such as mmWave radar and depth camera. The question is, can we support in-air gesture control on legacy devices without any hardware modifications?

In-air gesture control extends a touch screen and enables contact-less interaction, thus has become a popular research direction in the past few years. Prior work has implemented this functionality based on cameras, acoustic signals, and Wi-Fi via existing hardware on commercial devices. However, these methods have low user acceptance. Solutions based on cameras and acoustic signals raise privacy concerns, while WiFi-based solutions are vulnerable to background noise. As a result, these methods are not commercialized and recent flagship smartphones have implemented in-air gesture recognition by adding extra hardware on-board, such as mmWave radar and depth camera. The question is, can we support in-air gesture control on legacy devices without any hardware modifications?

Training process is crucial for the deployment of the network in applications which have two strict requirements on both accuracy and robustness. However, most existing approaches are in a dilemma, i.e. model accuracy and robustness form an embarrassing tradeoff - the improvement of one leads to the drop of the other. The challenge remains as for we try to improve the accuracy and robustness simultaneously. In this paper, we propose a novel training method via introducing the auxiliary classifiers for training on corrupted samples, while the clean samples are normally trained with the primary classifier. In the training stage, a novel distillation method named input-aware self distillation is proposed to facilitate the primary classifier to learn the robust information from auxiliary classifiers. Along with it, a new normalization method - selective batch normalization is proposed to prevent the model from the negative influence of corrupted images. At the end of training period, a L2-norm penalty is applied to the weights of primary and auxiliary classifiers such that their weights are asymptotically identical. In the stage of inference, only the primary classifier is used and thus no extra computation and storage are needed. Extensive experiments on CIFAR10, CIFAR100 and ImageNet show that noticeable improvements on both accuracy and robustness can be observed by the proposed auxiliary training. On average, auxiliary training achieves 2.21% accuracy and 21.64% robustness (measured by corruption error) improvements over traditional training methods on CIFAR100. Codes has been released on github.

Feature distillation, a primary method in knowledge distillation, always leads to significant accuracy improvements. Most existing methods distill features in the teacher network through a manually designed transformation. In this paper, we propose a novel distillation method named task-oriented feature distillation (TOFD) where the transformation is convolutional layers that are trained in a data-driven manner by task loss. As a result, the task-oriented information in the features can be captured and distilled to students. Moreover, an orthogonal loss is applied to the feature resizing layer in TOFD to improve the performance of knowledge distillation. Experiments show that TOFD outperforms other distillation methods by a large margin on both image classification and 3D classification tasks. Codes have been released in Github.

Convolutional neural networks have been widely deployed in various application scenarios. In order to extend the applications' boundaries to some accuracy-crucial domains, researchers have been investigating approaches to boost accuracy through either deeper or wider network structures, which brings with them the exponential increment of the computational and storage cost, delaying the responding time. In this paper, we propose a general training framework named self distillation, which notably enhances the performance (accuracy) of convolutional neural networks through shrinking the size of the network rather than aggrandizing it. Different from traditional knowledge distillation - a knowledge transformation methodology among networks, which forces student neural networks to approximate the softmax layer outputs of pre-trained teacher neural networks, the proposed self distillation framework distills knowledge within network itself. The networks are firstly divided into several sections. Then the knowledge in the deeper portion of the networks is squeezed into the shallow ones. Experiments further prove the generalization of the proposed self distillation framework: enhancement of accuracy at average level is 2.65%, varying from 0.61% in ResNeXt as minimum to 4.07% in VGG19 as maximum. In addition, it can also provide flexibility of depth-wise scalable inference on resource-limited edge devices. Our codes have been released on github.

Remarkable achievements have been attained by deep neural networks in various applications. However, the increasing depth and width of such models also lead to explosive growth in both storage and computation, which has restricted the deployment of deep neural networks on resource-limited edge devices. To address this problem, we propose the so-called SCAN framework for networks training and inference, which is orthogonal and complementary to existing acceleration and compression methods. The proposed SCAN firstly divides neural networks into multiple sections according to their depth and constructs shallow classifiers upon the intermediate features of different sections. Moreover, attention modules and knowledge distillation are utilized to enhance the accuracy of shallow classifiers. Based on this architecture, we further propose a threshold controlled scalable inference mechanism to approach human-like sample-specific inference. Experimental results show that SCAN can be easily equipped on various neural networks without any adjustment on hyper-parameters or neural networks architectures, yielding significant performance gain on CIFAR100 and ImageNet. Codes will be released on github soon.

Video object segmentation is a critical first step in virtual advertisement system. In this paper, we present belief propagation based method to increase the accuracy and efficiency of segmentation. The background color models are built as a mixture of a global color model and a per-pixel color model, while foreground color likelihood is modeled by spatially global Gaussian Mixture, and a realistic model of shadow effects is built also. The neighborhood dependencies are built with a MRF prior. Belief propagation is used to compute the marginal probability in MRF, and the video objects are segmented effectively. Experiment results show that the algorithm has the good performance, and has been used in our virtual advertisement system.