Zhi Wang

Chen Tang, Kai Ouyang, Zhi Wang et al.

The exponentially large discrete search space in mixed-precision quantization (MPQ) makes it hard to determine the optimal bit-width for each layer. Previous works usually resort to iterative search methods on the training set, which consume hundreds or even thousands of GPU-hours. In this study, we reveal that some unique learnable parameters in quantization, namely the scale factors in the quantizer, can serve as importance indicators of a layer, reflecting the contribution of that layer to the final accuracy at certain bit-widths. These importance indicators naturally perceive the numerical transformation during quantization-aware training, which can precisely provide quantization sensitivity metrics of layers. However, a deep network always contains hundreds of such indicators, and training them one by one would lead to an excessive time cost. To overcome this issue, we propose a joint training scheme that can obtain all indicators at once. It considerably speeds up the indicators training process by parallelizing the original sequential training processes. With these learned importance indicators, we formulate the MPQ search problem as a one-time integer linear programming (ILP) problem. That avoids the iterative search and significantly reduces search time without limiting the bit-width search space. For example, MPQ search on ResNet18 with our indicators takes only 0.06 s, which improves time efficiency exponentially compared to iterative search methods. Also, extensive experiments show our approach can achieve SOTA accuracy on ImageNet for far-ranging models with various constraints (e.g., BitOps, compress rate). Code is available on https://github.com/1hunters/LIMPQ.

Hang Guo, Tao Dai, Mingyan Zhu et al.

Recognizing characters from low-resolution (LR) text images poses a significant challenge due to the information deficiency as well as the noise and blur in low-quality images. Current solutions for low-resolution text recognition (LTR) typically rely on a two-stage pipeline that involves super-resolution as the first stage followed by the second-stage recognition. Although this pipeline is straightforward and intuitive, it has to use an additional super-resolution network, which causes inefficiencies during training and testing. Moreover, the recognition accuracy of the second stage heavily depends on the reconstruction quality of the first stage, causing ineffectiveness. In this work, we attempt to address these challenges from a novel perspective: adapting the recognizer to low-resolution inputs by transferring the knowledge from the high-resolution. Guided by this idea, we propose an efficient and effective knowledge distillation framework to achieve multi-level knowledge transfer. Specifically, the visual focus loss is proposed to extract the character position knowledge with resolution gap reduction and character region focus, the semantic contrastive loss is employed to exploit the contextual semantic knowledge with contrastive learning, and the soft logits loss facilitates both local word-level and global sequence-level learning from the soft teacher label. Extensive experiments show that the proposed one-stage pipeline significantly outperforms super-resolution based two-stage frameworks in terms of effectiveness and efficiency, accompanied by favorable robustness. Code is available at https://github.com/csguoh/KD-LTR.

Chen Tang, Li Lyna Zhang, Huiqiang Jiang et al. · microsoft-research

Neural Architecture Search (NAS) has shown promising performance in the automatic design of vision transformers (ViT) exceeding 1G FLOPs. However, designing lightweight and low-latency ViT models for diverse mobile devices remains a big challenge. In this work, we propose ElasticViT, a two-stage NAS approach that trains a high-quality ViT supernet over a very large search space that supports a wide range of mobile devices, and then searches an optimal sub-network (subnet) for direct deployment. However, prior supernet training methods that rely on uniform sampling suffer from the gradient conflict issue: the sampled subnets can have vastly different model sizes (e.g., 50M vs. 2G FLOPs), leading to different optimization directions and inferior performance. To address this challenge, we propose two novel sampling techniques: complexity-aware sampling and performance-aware sampling. Complexity-aware sampling limits the FLOPs difference among the subnets sampled across adjacent training steps, while covering different-sized subnets in the search space. Performance-aware sampling further selects subnets that have good accuracy, which can reduce gradient conflicts and improve supernet quality. Our discovered models, ElasticViT models, achieve top-1 accuracy from 67.2% to 80.0% on ImageNet from 60M to 800M FLOPs without extra retraining, outperforming all prior CNNs and ViTs in terms of accuracy and latency. Our tiny and small models are also the first ViT models that surpass state-of-the-art CNNs with significantly lower latency on mobile devices. For instance, ElasticViT-S1 runs 2.62x faster than EfficientNet-B0 with 0.1% higher accuracy.

Mingqing Wang, Jiawei Li, Zhenyang Li et al.

Unsupervised anomaly detection (UAD) has been widely implemented in industrial and medical applications, which reduces the cost of manual annotation and improves efficiency in disease diagnosis. Recently, deep auto-encoder with its variants has demonstrated its advantages in many UAD scenarios. Training on the normal data, these models are expected to locate anomalies by producing higher reconstruction error for the abnormal areas than the normal ones. However, this assumption does not always hold because of the uncontrollable generalization capability. To solve this problem, we present LSGS, a method that builds on Vector Quantised-Variational Autoencoder (VQVAE) with a novel aggregated codebook and transformers with global attention. In this work, the VQVAE focus on feature extraction and reconstruction of images, and the transformers fit the manifold and locate anomalies in the latent space. Then, leveraging the generated encoding sequences that conform to a normal distribution, we can reconstruct a more accurate image for locating the anomalies. Experiments on various datasets demonstrate the effectiveness of the proposed method.

Zhiwei Hao, Yong Luo, Zhi Wang et al.

Recently, the compression and deployment of powerful deep neural networks (DNNs) on resource-limited edge devices to provide intelligent services have become attractive tasks. Although knowledge distillation (KD) is a feasible solution for compression, its requirement on the original dataset raises privacy concerns. In addition, it is common to integrate multiple pretrained models to achieve satisfactory performance. How to compress multiple models into a tiny model is challenging, especially when the original data are unavailable. To tackle this challenge, we propose a framework termed collaborative data-free knowledge distillation via multi-level feature sharing (CDFKD-MFS), which consists of a multi-header student module, an asymmetric adversarial data-free KD module, and an attention-based aggregation module. In this framework, the student model equipped with a multi-level feature-sharing structure learns from multiple teacher models and is trained together with a generator in an asymmetric adversarial manner. When some real samples are available, the attention module adaptively aggregates predictions of the student headers, which can further improve performance. We conduct extensive experiments on three popular computer visual datasets. In particular, compared with the most competitive alternative, the accuracy of the proposed framework is 1.18\% higher on the CIFAR-100 dataset, 1.67\% higher on the Caltech-101 dataset, and 2.99\% higher on the mini-ImageNet dataset.

Xiaohan Lan, Yitian Yuan, Xin Wang et al.

Temporal Sentence Grounding in Videos (TSGV), which aims to ground a natural language sentence in an untrimmed video, has drawn widespread attention over the past few years. However, recent studies have found that current benchmark datasets may have obvious moment annotation biases, enabling several simple baselines even without training to achieve SOTA performance. In this paper, we take a closer look at existing evaluation protocols, and find both the prevailing dataset and evaluation metrics are the devils that lead to untrustworthy benchmarking. Therefore, we propose to re-organize the two widely-used datasets, making the ground-truth moment distributions different in the training and test splits, i.e., out-of-distribution (OOD) test. Meanwhile, we introduce a new evaluation metric "dR@n,IoU@m" that discounts the basic recall scores to alleviate the inflating evaluation caused by biased datasets. New benchmarking results indicate that our proposed evaluation protocols can better monitor the research progress. Furthermore, we propose a novel causality-based Multi-branch Deconfounding Debiasing (MDD) framework for unbiased moment prediction. Specifically, we design a multi-branch deconfounder to eliminate the effects caused by multiple confounders with causal intervention. In order to help the model better align the semantics between sentence queries and video moments, we enhance the representations during feature encoding. Specifically, for textual information, the query is parsed into several verb-centered phrases to obtain a more fine-grained textual feature. For visual information, the positional information has been decomposed from moment features to enhance representations of moments with diverse locations. Extensive experiments demonstrate that our proposed approach can achieve competitive results among existing SOTA approaches and outperform the base model with great gains.

Jinmei Liu, Zhi Wang, Chunlin Chen et al.

Bayesian policy reuse (BPR) is a general policy transfer framework for selecting a source policy from an offline library by inferring the task belief based on some observation signals and a trained observation model. In this paper, we propose an improved BPR method to achieve more efficient policy transfer in deep reinforcement learning (DRL). First, most BPR algorithms use the episodic return as the observation signal that contains limited information and cannot be obtained until the end of an episode. Instead, we employ the state transition sample, which is informative and instantaneous, as the observation signal for faster and more accurate task inference. Second, BPR algorithms usually require numerous samples to estimate the probability distribution of the tabular-based observation model, which may be expensive and even infeasible to learn and maintain, especially when using the state transition sample as the signal. Hence, we propose a scalable observation model based on fitting state transition functions of source tasks from only a small number of samples, which can generalize to any signals observed in the target task. Moreover, we extend the offline-mode BPR to the continual learning setting by expanding the scalable observation model in a plug-and-play fashion, which can avoid negative transfer when faced with new unknown tasks. Experimental results show that our method can consistently facilitate faster and more efficient policy transfer.

Chenghao Gu, Haolan Kang, Junchao Lin et al.

The rise of generalist robotic policies has created an exponential demand for large-scale training data. However, on-robot data collection is labor-intensive and often limited to specific environments. In contrast, open-world images capture a vast diversity of real-world scenes that naturally align with robotic manipulation tasks, offering a promising avenue for low-cost, large-scale robot data acquisition. Despite this potential, the lack of associated robot actions hinders the practical use of open-world images for robot learning, leaving this rich visual resource largely unexploited. To bridge this gap, we propose IGen, a framework that scalably generates realistic visual observations and executable actions from open-world images. IGen first converts unstructured 2D pixels into structured 3D scene representations suitable for scene understanding and manipulation. It then leverages the reasoning capabilities of vision-language models to transform scene-specific task instructions into high-level plans and generate low-level actions as SE(3) end-effector pose sequences. From these poses, it synthesizes dynamic scene evolution and renders temporally coherent visual observations. Experiments validate the high quality of visuomotor data generated by IGen, and show that policies trained solely on IGen-synthesized data achieve performance comparable to those trained on real-world data. This highlights the potential of IGen to support scalable data generation from open-world images for generalist robotic policy training.

Yiwen Shan, Dong Hu, Zhi Wang

Due to the high flexibility and remarkable performance, low-rank approximation methods has been widely studied for color image denoising. However, those methods mostly ignore either the cross-channel difference or the spatial variation of noise, which limits their capacity in real world color image denoising. To overcome those drawbacks, this paper is proposed to denoise color images with a double-weighted truncated nuclear norm minus truncated Frobenius norm minimization (DtNFM) method. Through exploiting the nonlocal self-similarity of the noisy image, the similar structures are gathered and a series of similar patch matrices are constructed. For each group, the DtNFM model is conducted for estimating its denoised version. The denoised image would be obtained by concatenating all the denoised patch matrices. The proposed DtNFM model has two merits. First, it models and utilizes both the cross-channel difference and the spatial variation of noise. This provides sufficient flexibility for handling the complex distribution of noise in real world images. Second, the proposed DtNFM model provides a close approximation to the underlying clean matrix since it can treat different rank components flexibly. To solve the problem resulted from DtNFM model, an accurate and effective algorithm is proposed by exploiting the framework of the alternating direction method of multipliers (ADMM). The generated subproblems are discussed in detail. And their global optima can be easily obtained in closed-form. Rigorous mathematical derivation proves that the solution sequences generated by the algorithm converge to a single critical point. Extensive experiments on synthetic and real noise datasets demonstrate that the proposed method outperforms many state-of-the-art color image denoising methods.

Chen Tang, Yuan Meng, Jiacheng Jiang et al.

Quantization is of significance for compressing the over-parameterized deep neural models and deploying them on resource-limited devices. Fixed-precision quantization suffers from performance drop due to the limited numerical representation ability. Conversely, mixed-precision quantization (MPQ) is advocated to compress the model effectively by allocating heterogeneous bit-width for layers. MPQ is typically organized into a searching-retraining two-stage process. In this paper, we devise a one-shot training-searching paradigm for mixed-precision model compression. Specifically, in the first stage, all potential bit-width configurations are coupled and thus optimized simultaneously within a set of shared weights. However, our observations reveal a previously unseen and severe bit-width interference phenomenon among highly coupled weights during optimization, leading to considerable performance degradation under a high compression ratio. To tackle this problem, we first design a bit-width scheduler to dynamically freeze the most turbulent bit-width of layers during training, to ensure the rest bit-widths converged properly. Then, taking inspiration from information theory, we present an information distortion mitigation technique to align the behavior of the bad-performing bit-widths to the well-performing ones. In the second stage, an inference-only greedy search scheme is devised to evaluate the goodness of configurations without introducing any additional training costs. Extensive experiments on three representative models and three datasets demonstrate the effectiveness of the proposed method. Code can be available on \href{https://www.github.com/1hunters/retraining-free-quantization}{https://github.com/1hunters/retraining-free-quantization}.

Jinmei Liu, Wenbin Li, Xiangyu Yue et al.

We study continual offline reinforcement learning, a practical paradigm that facilitates forward transfer and mitigates catastrophic forgetting to tackle sequential offline tasks. We propose a dual generative replay framework that retains previous knowledge by concurrent replay of generated pseudo-data. First, we decouple the continual learning policy into a diffusion-based generative behavior model and a multi-head action evaluation model, allowing the policy to inherit distributional expressivity for encompassing a progressive range of diverse behaviors. Second, we train a task-conditioned diffusion model to mimic state distributions of past tasks. Generated states are paired with corresponding responses from the behavior generator to represent old tasks with high-fidelity replayed samples. Finally, by interleaving pseudo samples with real ones of the new task, we continually update the state and behavior generators to model progressively diverse behaviors, and regularize the multi-head critic via behavior cloning to mitigate forgetting. Experiments demonstrate that our method achieves better forward transfer with less forgetting, and closely approximates the results of using previous ground-truth data due to its high-fidelity replay of the sample space. Our code is available at \href{https://github.com/NJU-RL/CuGRO}{https://github.com/NJU-RL/CuGRO}.

Zijie Fang, Yifeng Wang, Peizhang Xie et al.

Tissue semantic segmentation is one of the key tasks in computational pathology. To avoid the expensive and laborious acquisition of pixel-level annotations, a wide range of studies attempt to adopt the class activation map (CAM), a weakly-supervised learning scheme, to achieve pixel-level tissue segmentation. However, CAM-based methods are prone to suffer from under-activation and over-activation issues, leading to poor segmentation performance. To address this problem, we propose a novel weakly-supervised semantic segmentation framework for histopathological images based on image-mixing synthesis and consistency regularization, dubbed HisynSeg. Specifically, synthesized histopathological images with pixel-level masks are generated for fully-supervised model training, where two synthesis strategies are proposed based on Mosaic transformation and Bézier mask generation. Besides, an image filtering module is developed to guarantee the authenticity of the synthesized images. In order to further avoid the model overfitting to the occasional synthesis artifacts, we additionally propose a novel self-supervised consistency regularization, which enables the real images without segmentation masks to supervise the training of the segmentation model. By integrating the proposed techniques, the HisynSeg framework successfully transforms the weakly-supervised semantic segmentation problem into a fully-supervised one, greatly improving the segmentation accuracy. Experimental results on three datasets prove that the proposed method achieves a state-of-the-art performance. Code is available at https://github.com/Vison307/HisynSeg.

Shilin Zhang, Zican Hu, Wenhao Wu et al.

Offline meta-RL usually tackles generalization by inferring task beliefs from high-quality samples or warmup explorations. The restricted form limits their generality and usability since these supervision signals are expensive and even infeasible to acquire in advance for unseen tasks. Learning directly from the raw text about decision tasks is a promising alternative to leverage a much broader source of supervision. In the paper, we propose \textbf{T}ext-to-\textbf{D}ecision \textbf{A}gent (\textbf{T2DA}), a simple and scalable framework that supervises offline meta-RL with natural language. We first introduce a generalized world model to encode multi-task decision data into a dynamics-aware embedding space. Then, inspired by CLIP, we predict which textual description goes with which decision embedding, effectively bridging their semantic gap via contrastive language-decision pre-training and aligning the text embeddings to comprehend the environment dynamics. After training the text-conditioned generalist policy, the agent can directly realize zero-shot text-to-decision generation in response to language instructions. Comprehensive experiments on MuJoCo and Meta-World benchmarks show that T2DA facilitates high-capacity zero-shot generalization and outperforms various types of baselines. Our code is available at https://github.com/NJU-RL/T2DA.

Yijun Liu, Yuan Meng, Fang Wu et al.

Large language models (LLMs) have exhibited exciting progress in multiple scenarios, while the huge computational demands hinder their deployments in lots of real-world applications. As an effective means to reduce memory footprint and inference cost, quantization also faces challenges in performance degradation at low bit-widths. Understanding the impact of quantization on LLM capabilities, especially the generalization ability, is crucial. However, the community's main focus remains on the algorithms and models of quantization, with insufficient attention given to whether the quantized models can retain the strong generalization abilities of LLMs. In this work, we fill this gap by providing a comprehensive benchmark suite for this research topic, including an evaluation system, detailed analyses, and a general toolbox. Specifically, based on the dominant pipeline in LLM quantization, we primarily explore the impact of calibration data distribution on the generalization of quantized LLMs and conduct the benchmark using more than 40 datasets within two main scenarios. Based on this benchmark, we conduct extensive experiments with two well-known LLMs (English and Chinese) and four quantization algorithms to investigate this topic in-depth, yielding several counter-intuitive and valuable findings, e.g., models quantized using a calibration set with the same distribution as the test data are not necessarily optimal. Besides, to facilitate future research, we also release a modular-designed toolbox, which decouples the overall pipeline into several separate components, e.g., base LLM module, dataset module, quantizer module, etc. and allows subsequent researchers to easily assemble their methods through a simple configuration. Our benchmark suite is publicly available at https://github.com/TsingmaoAI/MI-optimize

Alexander I. Cowen-Rivers, Wenlong Lyu, Rasul Tutunov et al.

In this work we rigorously analyse assumptions inherent to black-box optimisation hyper-parameter tuning tasks. Our results on the Bayesmark benchmark indicate that heteroscedasticity and non-stationarity pose significant challenges for black-box optimisers. Based on these findings, we propose a Heteroscedastic and Evolutionary Bayesian Optimisation solver (HEBO). HEBO performs non-linear input and output warping, admits exact marginal log-likelihood optimisation and is robust to the values of learned parameters. We demonstrate HEBO's empirical efficacy on the NeurIPS 2020 Black-Box Optimisation challenge, where HEBO placed first. Upon further analysis, we observe that HEBO significantly outperforms existing black-box optimisers on 108 machine learning hyperparameter tuning tasks comprising the Bayesmark benchmark. Our findings indicate that the majority of hyper-parameter tuning tasks exhibit heteroscedasticity and non-stationarity, multi-objective acquisition ensembles with Pareto front solutions improve queried configurations, and robust acquisition maximisers afford empirical advantages relative to their non-robust counterparts. We hope these findings may serve as guiding principles for practitioners of Bayesian optimisation. All code is made available at https://github.com/huawei-noah/HEBO.

Fang Hu, Xuexue Sun, Ke Qing et al.

Although deep learning (DL) shows powerful potential in cell segmentation tasks, it suffers from poor generalization as DL-based methods originally simplified cell segmentation in detecting cell membrane boundary, lacking prominent cellular structures to position overall differentiating. Moreover, the scarcity of annotated cell images limits the performance of DL models. Segmentation limitations of a single category of cell make massive practice difficult, much less, with varied modalities. In this paper, we introduce a novel semi-supervised cell segmentation method called Multi-Microscopic-view Cell semi-supervised Segmentation (MMCS), which can train cell segmentation models utilizing less labeled multi-posture cell images with different microscopy well. Technically, MMCS consists of Nucleus-assisted global recognition, Self-adaptive diameter filter, and Temporal-ensembling models. Nucleus-assisted global recognition adds additional cell nucleus channel to improve the global distinguishing performance of fuzzy cell membrane boundaries even when cells aggregate. Besides, self-adapted cell diameter filter can help separate multi-resolution cells with different morphology properly. It further leverages the temporal-ensembling models to improve the semi-supervised training process, achieving effective training with less labeled data. Additionally, optimizing the weight of unlabeled loss contributed to total loss also improve the model performance. Evaluated on the Tuning Set of NeurIPS 2022 Cell Segmentation Challenge (NeurIPS CellSeg), MMCS achieves an F1-score of 0.8239 and the running time for all cases is within the time tolerance.

Zeming Chen, Wenwei Zhang, Xinjiang Wang et al.

While the pseudo-label method has demonstrated considerable success in semi-supervised object detection tasks, this paper uncovers notable limitations within this approach. Specifically, the pseudo-label method tends to amplify the inherent strengths of the detector while accentuating its weaknesses, which is manifested in the missed detection of pseudo-labels, particularly for small and tail category objects. To overcome these challenges, this paper proposes Mixed Pseudo Labels (MixPL), consisting of Mixup and Mosaic for pseudo-labeled data, to mitigate the negative impact of missed detections and balance the model's learning across different object scales. Additionally, the model's detection performance on tail categories is improved by resampling labeled data with relevant instances. Notably, MixPL consistently improves the performance of various detectors and obtains new state-of-the-art results with Faster R-CNN, FCOS, and DINO on COCO-Standard and COCO-Full benchmarks. Furthermore, MixPL also exhibits good scalability on large models, improving DINO Swin-L by 2.5% mAP and achieving nontrivial new records (60.2% mAP) on the COCO val2017 benchmark without extra annotations.

Jianghui Zhang, Yuanyuan Wang, Lina Guo et al.

JPEG is still the most widely used image compression algorithm. Most image compression algorithms only consider uncompressed original image, while ignoring a large number of already existing JPEG images. Recently, JPEG recompression approaches have been proposed to further reduce the size of JPEG files. However, those methods only consider JPEG lossless recompression, which is just a special case of the rate-distortion theorem. In this paper, we propose a unified lossly and lossless JPEG recompression framework, which consists of learned quantization table and Markovian hierarchical variational autoencoders. Experiments show that our method can achieve arbitrarily low distortion when the bitrate is close to the upper bound, namely the bitrate of the lossless compression model. To the best of our knowledge, this is the first learned method that bridges the gap between lossy and lossless recompression of JPEG images.

Weixiang Zhang, Shuzhao Xie, Chengwei Ren et al.

We propose EVOlutionary Selector (EVOS), an efficient training paradigm for accelerating Implicit Neural Representation (INR). Unlike conventional INR training that feeds all samples through the neural network in each iteration, our approach restricts training to strategically selected points, reducing computational overhead by eliminating redundant forward passes. Specifically, we treat each sample as an individual in an evolutionary process, where only those fittest ones survive and merit inclusion in training, adaptively evolving with the neural network dynamics. While this is conceptually similar to Evolutionary Algorithms, their distinct objectives (selection for acceleration vs. iterative solution optimization) require a fundamental redefinition of evolutionary mechanisms for our context. In response, we design sparse fitness evaluation, frequency-guided crossover, and augmented unbiased mutation to comprise EVOS. These components respectively guide sample selection with reduced computational cost, enhance performance through frequency-domain balance, and mitigate selection bias from cached evaluation. Extensive experiments demonstrate that our method achieves approximately 48%-66% reduction in training time while ensuring superior convergence without additional cost, establishing state-of-the-art acceleration among recent sampling-based strategies.

Mingzi Wang, Yuan Meng, Chen Tang et al.

The co-design of neural network architectures, quantization precisions, and hardware accelerators offers a promising approach to achieving an optimal balance between performance and efficiency, particularly for model deployment on resource-constrained edge devices. In this work, we propose the JAQ Framework, which jointly optimizes the three critical dimensions. However, effectively automating the design process across the vast search space of those three dimensions poses significant challenges, especially when pursuing extremely low-bit quantization. Specifical, the primary challenges include: (1) Memory overhead in software-side: Low-precision quantization-aware training can lead to significant memory usage due to storing large intermediate features and latent weights for back-propagation, potentially causing memory exhaustion. (2) Search time-consuming in hardware-side: The discrete nature of hardware parameters and the complex interplay between compiler optimizations and individual operators make the accelerator search time-consuming. To address these issues, JAQ mitigates the memory overhead through a channel-wise sparse quantization (CSQ) scheme, selectively applying quantization to the most sensitive components of the model during optimization. Additionally, JAQ designs BatchTile, which employs a hardware generation network to encode all possible tiling modes, thereby speeding up the search for the optimal compiler mapping strategy. Extensive experiments demonstrate the effectiveness of JAQ, achieving approximately 7% higher Top-1 accuracy on ImageNet compared to previous methods and reducing the hardware search time per iteration to 0.15 seconds.

Jiacheng Jiang, Yuan Meng, Chen Tang et al.

Research on loss surface geometry, such as Sharpness-Aware Minimization (SAM), shows that flatter minima improve generalization. Recent studies further reveal that flatter minima can also reduce the domain generalization (DG) gap. However, existing flatness-based DG techniques predominantly operate within a full-precision training process, which is impractical for deployment on resource-constrained edge devices that typically rely on lower bit-width representations (e.g., 4 bits, 3 bits). Consequently, low-precision quantization-aware training is critical for optimizing these techniques in real-world applications. In this paper, we observe a significant degradation in performance when applying state-of-the-art DG-SAM methods to quantized models, suggesting that current approaches fail to preserve generalizability during the low-precision training process. To address this limitation, we propose a novel Gradient-Adaptive Quantization-Aware Training (GAQAT) framework for DG. Our approach begins by identifying the scale-gradient conflict problem in low-precision quantization, where the task loss and smoothness loss induce conflicting gradients for the scaling factors of quantizers, with certain layers exhibiting opposing gradient directions. This conflict renders the optimization of quantized weights highly unstable. To mitigate this, we further introduce a mechanism to quantify gradient inconsistencies and selectively freeze the gradients of scaling factors, thereby stabilizing the training process and enhancing out-of-domain generalization. Extensive experiments validate the effectiveness of the proposed GAQAT framework. On PACS, our 3-bit and 4-bit models outperform direct DG-QAT integration by up to 4.5%. On DomainNet, the 4-bit model achieves near-lossless performance compared to full precision, with improvements of 1.39% (4-bit) and 1.06% (3-bit) over the SOTA QAT baseline.

Jiacheng Jiang, Yuan Meng, Chen Tang et al.

Current quantization-aware training (QAT) methods primarily focus on enhancing the performance of quantized models on in-distribution (I.D) data, while overlooking the potential performance degradation on out-of-distribution (OOD) data. In this paper, we first substantiate this problem through rigorous experiment, showing that QAT can lead to a significant OOD generalization performance degradation. Further, we find the contradiction between the perspective that flatness of loss landscape gives rise to superior OOD generalization and the phenomenon that QAT lead to a sharp loss landscape, can cause the above problem. Therefore, we propose a flatness-oriented QAT method, FQAT, to achieve generalizable QAT. Specifically, i) FQAT introduces a layer-wise freezing mechanism to mitigate the gradient conflict issue between dual optimization objectives (i.e., vanilla QAT and flatness). ii) FQAT proposes an disorder-guided adaptive freezing algorithm to dynamically determines which layers to freeze at each training step, effectively addressing the challenges caused by interference between layers. A gradient disorder metric is designed to help the algorithm identify unstable layers during training. Extensive experiments on influential OOD benchmark demonstrate the superiority of our method over state-of-the-art baselines under both I.D and OOD image classification tasks.

Taolin Zhang, Sunan He, Dai Tao et al.

In recent years, vision language pre-training frameworks have made significant progress in natural language processing and computer vision, achieving remarkable performance improvement on various downstream tasks. However, when extended to point cloud data, existing works mainly focus on building task-specific models, and fail to extract universal 3D vision-language embedding that generalize well. We carefully investigate three common tasks in semantic 3D scene understanding, and derive key insights into the development of a pre-training model. Motivated by these observations, we propose a vision-language pre-training framework 3DVLP (3D vision-language pre-training with object contrastive learning), which transfers flexibly on 3D vision-language downstream tasks. 3DVLP takes visual grounding as the proxy task and introduces Object-level IoU-guided Detection (OID) loss to obtain high-quality proposals in the scene. Moreover, we design Object-level Cross-Contrastive alignment (OCC) task and Object-level Self-Contrastive learning (OSC) task to align the objects with descriptions and distinguish different objects in the scene, respectively. Extensive experiments verify the excellent performance of 3DVLP on three 3D vision-language tasks, reflecting its superiority in semantic 3D scene understanding.

Yuan Wang, Wei Zhuo, Yucong Li et al.

In this work, we present a fully self-supervised framework for semantic segmentation(FS^4). A fully bootstrapped strategy for semantic segmentation, which saves efforts for the huge amount of annotation, is crucial for building customized models from end-to-end for open-world domains. This application is eagerly needed in realistic scenarios. Even though recent self-supervised semantic segmentation methods have gained great progress, these works however heavily depend on the fully-supervised pretrained model and make it impossible a fully self-supervised pipeline. To solve this problem, we proposed a bootstrapped training scheme for semantic segmentation, which fully leveraged the global semantic knowledge for self-supervision with our proposed PGG strategy and CAE module. In particular, we perform pixel clustering and assignments for segmentation supervision. Preventing it from clustering a mess, we proposed 1) a pyramid-global-guided (PGG) training strategy to supervise the learning with pyramid image/patch-level pseudo labels, which are generated by grouping the unsupervised features. The stable global and pyramid semantic pseudo labels can prevent the segmentation from learning too many clutter regions or degrading to one background region; 2) in addition, we proposed context-aware embedding (CAE) module to generate global feature embedding in view of its neighbors close both in space and appearance in a non-trivial way. We evaluate our method on the large-scale COCO-Stuff dataset and achieved 7.19 mIoU improvements on both things and stuff objects

Cheng Chen, Yichun Yin, Lifeng Shang et al.

In recent years, researchers tend to pre-train ever-larger language models to explore the upper limit of deep models. However, large language model pre-training costs intensive computational resources and most of the models are trained from scratch without reusing the existing pre-trained models, which is wasteful. In this paper, we propose bert2BERT, which can effectively transfer the knowledge of an existing smaller pre-trained model (e.g., BERT_BASE) to a large model (e.g., BERT_LARGE) through parameter initialization and significantly improve the pre-training efficiency of the large model. Specifically, we extend the previous function-preserving on Transformer-based language model, and further improve it by proposing advanced knowledge for large model's initialization. In addition, a two-stage pre-training method is proposed to further accelerate the training process. We did extensive experiments on representative PLMs (e.g., BERT and GPT) and demonstrate that (1) our method can save a significant amount of training cost compared with baselines including learning from scratch, StackBERT and MSLT; (2) our method is generic and applicable to different types of pre-trained models. In particular, bert2BERT saves about 45% and 47% computational cost of pre-training BERT_BASE and GPT_BASE by reusing the models of almost their half sizes. The source code will be publicly available upon publication.

Xiaohan Lan, Yitian Yuan, Xin Wang et al.

Temporal sentence grounding in videos(TSGV), which aims to localize one target segment from an untrimmed video with respect to a given sentence query, has drawn increasing attentions in the research community over the past few years. Different from the task of temporal action localization, TSGV is more flexible since it can locate complicated activities via natural languages, without restrictions from predefined action categories. Meanwhile, TSGV is more challenging since it requires both textual and visual understanding for semantic alignment between two modalities(i.e., text and video). In this survey, we give a comprehensive overview for TSGV, which i) summarizes the taxonomy of existing methods, ii) provides a detailed description of the evaluation protocols(i.e., datasets and metrics) to be used in TSGV, and iii) in-depth discusses potential problems of current benchmarking designs and research directions for further investigations. To the best of our knowledge, this is the first systematic survey on temporal sentence grounding. More specifically, we first discuss existing TSGV approaches by grouping them into four categories, i.e., two-stage methods, end-to-end methods, reinforcement learning-based methods, and weakly supervised methods. Then we present the benchmark datasets and evaluation metrics to assess current research progress. Finally, we discuss some limitations in TSGV through pointing out potential problems improperly resolved in the current evaluation protocols, which may push forwards more cutting edge research in TSGV. Besides, we also share our insights on several promising directions, including three typical tasks with new and practical settings based on TSGV.

Ziyun Zeng, Jinpeng Wang, Bin Chen et al.

Deep hashing approaches, including deep quantization and deep binary hashing, have become a common solution to large-scale image retrieval due to their high computation and storage efficiency. Most existing hashing methods cannot produce satisfactory results for fine-grained retrieval, because they usually adopt the outputs of the last CNN layer to generate binary codes. Since deeper layers tend to summarize visual clues, e.g., texture, into abstract semantics, e.g., dogs and cats, the feature produced by the last CNN layer is less effective in capturing subtle but discriminative visual details that mostly exist in shallow layers. To improve fine-grained image hashing, we propose Pyramid Hybrid Pooling Quantization (PHPQ). Specifically, we propose a Pyramid Hybrid Pooling (PHP) module to capture and preserve fine-grained semantic information from multi-level features, which emphasizes the subtle discrimination of different sub-categories. Besides, we propose a learnable quantization module with a partial codebook attention mechanism, which helps to optimize the most relevant codewords and improves the quantization. Comprehensive experiments on two widely-used public benchmarks, i.e., CUB-200-2011 and Stanford Dogs, demonstrate that PHPQ outperforms state-of-the-art methods.

Shiji Zhou, Han Zhao, Shanghang Zhang et al.

Models trained with offline data often suffer from continual distribution shifts and expensive labeling in changing environments. This calls for a new online learning paradigm where the learner can continually adapt to changing environments with limited labels. In this paper, we propose a new online setting -- Online Active Continual Adaptation, where the learner aims to continually adapt to changing distributions using both unlabeled samples and active queries of limited labels. To this end, we propose Online Self-Adaptive Mirror Descent (OSAMD), which adopts an online teacher-student structure to enable online self-training from unlabeled data, and a margin-based criterion that decides whether to query the labels to track changing distributions. Theoretically, we show that, in the separable case, OSAMD has an $O({T}^{2/3})$ dynamic regret bound under mild assumptions, which is aligned with the $Ω(T^{2/3})$ lower bound of online learning algorithms with full labels. In the general case, we show a regret bound of $O({T}^{2/3} + α^* T)$, where $α^*$ denotes the separability of domains and is usually small. Our theoretical results show that OSAMD can fast adapt to changing environments with active queries. Empirically, we demonstrate that OSAMD achieves favorable regrets under changing environments with limited labels on both simulated and real-world data, which corroborates our theoretical findings.

Zhi Wang, Wenwu Zhu, Lifeng Sun et al.

In this paper, we investigate the recent studies on multimedia edge computing, from sensing not only traditional visual/audio data but also individuals' geographical preference and mobility behaviors, to performing distributed machine learning over such data using the joint edge and cloud infrastructure and using evolutional strategies like reinforcement learning and online learning at edge devices to optimize the quality of experience for multimedia services at the last mile proactively. We provide both a retrospective view of recent rapid migration (resp. merge) of cloud multimedia to (resp. and) edge-aware multimedia and insights on the fundamental guidelines for designing multimedia edge computing strategies that target satisfying the changing demand of quality of experience. By showing the recent research studies and industrial solutions, we also provide future directions towards high-quality multimedia services over edge computing.

Cheng Chen, Yichun Yin, Lifeng Shang et al.

Task-agnostic knowledge distillation, a teacher-student framework, has been proved effective for BERT compression. Although achieving promising results on NLP tasks, it requires enormous computational resources. In this paper, we propose Extract Then Distill (ETD), a generic and flexible strategy to reuse the teacher's parameters for efficient and effective task-agnostic distillation, which can be applied to students of any size. Specifically, we introduce two variants of ETD, ETD-Rand and ETD-Impt, which extract the teacher's parameters in a random manner and by following an importance metric respectively. In this way, the student has already acquired some knowledge at the beginning of the distillation process, which makes the distillation process converge faster. We demonstrate the effectiveness of ETD on the GLUE benchmark and SQuAD. The experimental results show that: (1) compared with the baseline without an ETD strategy, ETD can save 70\% of computation cost. Moreover, it achieves better results than the baseline when using the same computing resource. (2) ETD is generic and has been proven effective for different distillation methods (e.g., TinyBERT and MiniLM) and students of different sizes. The source code will be publicly available upon publication.

Chenghao Hu, Jingyan Jiang, Zhi Wang

The emerging concern about data privacy and security has motivated the proposal of federated learning, which allows nodes to only synchronize the locally-trained models instead their own original data. Conventional federated learning architecture, inherited from the parameter server design, relies on highly centralized topologies and the assumption of large nodes-to-server bandwidths. However, in real-world federated learning scenarios the network capacities between nodes are highly uniformly distributed and smaller than that in a datacenter. It is of great challenges for conventional federated learning approaches to efficiently utilize network capacities between nodes. In this paper, we propose a model segment level decentralized federated learning to tackle this problem. In particular, we propose a segmented gossip approach, which not only makes full utilization of node-to-node bandwidth, but also has good training convergence. The experimental results show that even the training time can be highly reduced as compared to centralized federated learning.

Xueying Tang, Zhi Wang, Jingchen Liu et al.

Computer simulations have become a popular tool of assessing complex skills such as problem-solving skills. Log files of computer-based items record the entire human-computer interactive processes for each respondent. The response processes are very diverse, noisy, and of nonstandard formats. Few generic methods have been developed for exploiting the information contained in process data. In this article, we propose a method to extract latent variables from process data. The method utilizes a sequence-to-sequence autoencoder to compress response processes into standard numerical vectors. It does not require prior knowledge of the specific items and human-computers interaction patterns. The proposed method is applied to both simulated and real process data to demonstrate that the resulting latent variables extract useful information from the response processes.

Ge Ma, Zhi Wang, Miao Zhang et al.

Today's Internet has witnessed an increase in the popularity of mobile video streaming, which is expected to exceed 3/4 of the global mobile data traffic by 2019. To satisfy the considerable amount of mobile video requests, video service providers have been pushing their content delivery infrastructure to edge networks--from regional CDN servers to peer CDN servers (e.g., smartrouters in users' homes)--to cache content and serve users with storage and network resources nearby. Among the edge network content caching paradigms, Wi-Fi access point caching and cellular base station caching have become two mainstream solutions. Thus, understanding the effectiveness and performance of these solutions for large-scale mobile video delivery is important. However, the characteristics and request patterns of mobile video streaming are unclear in practical wireless network. In this paper, we use real-world datasets containing 50 million trace items of nearly 2 million users viewing more than 0.3 million unique videos using mobile devices in a metropolis in China over 2 weeks, not only to understand the request patterns and user behaviors in mobile video streaming, but also to evaluate the effectiveness of Wi-Fi and cellular-based edge content caching solutions. To understand performance of edge content caching for mobile video streaming, we first present temporal and spatial video request patterns, and we analyze their impacts on caching performance using frequency-domain and entropy analysis approaches. We then study the behaviors of mobile video users, including their mobility and geographical migration behaviors. Using trace-driven experiments, we compare strategies for edge content caching including LRU and LFU, in terms of supporting mobile video requests. Moreover, we design an efficient caching strategy based on the measurement insights and experimentally evaluate its performance.

Ming Ma, Zhi Wang, Yankai Zhang et al.

Multihoming for a video Content Delivery Network (CDN) allows edge peering servers to deliver video chunks through different Internet Service Providers (ISPs), to achieve an improved quality of service (QoS) for video streaming users. However, since traditional strategies for a multihoming video CDN are simply designed according to static rules, e.g., simply sending traffic via a ISP which is the same as the ISP of client, they fail to dynamically allocate resources among different ISPs over time. In this paper, we perform measurement studies to demonstrate that such static allocation mechanism is inefficient to make full utilization of multiple ISPs' resources. To address this problem, we propose a dynamic flow scheduling strategy for multihoming video CDN. The challenge is to find the control parameters that can guide the ISP selection when performing flow scheduling. Using a data-driven approach, we find factors that have a major impact on the performance improvement in the dynamic flow scheduling. We further utilize an information gain approach to generate parameter combinations that can be used to guide the flow scheduling, i.e., to determine the ISP each request should be responded by. Our evaluation results demonstrate that our design effectively performs the flow scheduling. In particular, our design yields near optimal performance in a simulation of real-world multihoming setup.

Wen Hu, Zhi Wang, Lifeng Sun

Dynamic Adaptive Streaming over HTTP (DASH) has emerged as an increasingly popular paradigm for video streaming [13], in which a video is segmented into many chunks delivered to users by HTTP request/response over Transmission Control Protocol (TCP) con- nections. Therefore, it is intriguing to study the performance of strategies implemented in conventional TCPs, which are not dedicated for video streaming, e.g., whether chunks are efficiently delivered when users per- form interactions with the video players. In this paper, we conduct mea- surement studies on users chunk requesting traces in DASH from a rep- resentative video streaming provider, to investigate users behaviors in DASH, and TCP-connection-level traces from CDN servers, to investi- gate the performance of TCP for DASH. By studying how video chunks are delivered in both the slow start and congestion avoidance phases, our observations have revealed the performance characteristics of TCP for DASH as follows: (1) Request patterns in DASH have a great impact on the performance of TCP variations including cubic; (2) Strategies in conventional TCPs may cause user perceived quality degradation in DASH streaming; (3) Potential improvement to TCP strategies for better delivery in DASH can be further explored.

Zhi Wang, Lifeng Sun, Miao Zhang et al.

Mobile online social network services have seen a rapid increase, in which the huge amount of user-generated social media contents propagating between users via social connections has significantly challenged the traditional content delivery paradigm: First, replicating all of the contents generated by users to edge servers that well "fit" the receivers becomes difficult due to the limited bandwidth and storage capacities. Motivated by device-to-device (D2D) communication that allows users with smart devices to transfer content directly, we propose replicating bandwidth-intensive social contents in a device-to-device manner. Based on large-scale measurement studies on social content propagation and user mobility patterns in edge-network regions, we observe that (1) Device-to-device replication can significantly help users download social contents from nearby neighboring peers; (2) Both social propagation and mobility patterns affect how contents should be replicated; (3) The replication strategies depend on regional characteristics ({\em e.g.}, how users move across regions). Using these measurement insights, we propose a joint \emph{propagation- and mobility-aware} content replication strategy for edge-network regions, in which social contents are assigned to users in edge-network regions according to a joint consideration of social graph, content propagation and user mobility. We formulate the replication scheduling as an optimization problem and design distributed algorithm only using historical, local and partial information to solve it. Trace-driven experiments further verify the superiority of our proposal: compared with conventional pure movement-based and popularity-based approach, our design can significantly ($2-4$ times) improve the amount of social contents successfully delivered by device-to-device replication.

Chenglei Wu, Zhi Wang, Jiangchuan Liu et al.

Recent years have witnessed a dramatic increase of user-generated video services. In such user-generated video services, crowdsourced live streaming (e.g., Periscope, Twitch) has significantly challenged today's edge network infrastructure: today's edge networks (e.g., 4G, Wi-Fi) have limited uplink capacity support, making high-bitrate live streaming over such links fundamentally impossible. In this paper, we propose to let broadcasters (i.e., users who generate the video) upload crowdsourced video streams using aggregated network resources from multiple edge networks. There are several challenges in the proposal: First, how to design a framework that aggregates bandwidth from multiple edge networks? Second, how to make this framework transparent to today's crowdsourced live streaming services? Third, how to maximize the streaming quality for the whole system? We design a multi-objective and deployable bandwidth aggregation system BASS to address these challenges: (1) We propose an aggregation framework transparent to today's crowdsourced live streaming services, using an edge proxy box and aggregation cloud paradigm; (2) We dynamically allocate geo-distributed cloud aggregation servers to enable MPTCP (i.e., multi-path TCP), according to location and network characteristics of both broadcasters and the original streaming servers; (3) We maximize the overall performance gain for the whole system, by matching streams with the best aggregation paths.

Ming Ma, Zhi Wang, Ke Su et al.

Recent years have witnessed a new video delivery paradigm: smartrouter-based peer video content delivery network, which is enabled by smartrouters deployed at users' homes. ChinaCache (one of the largest CDN providers in China) and Youku (a video provider using smartrouters to assist video delivery) announced their cooperation in 2015, to create a new paradigm of content delivery based on householders' network resources. This new paradigm is different from the conventional peer-to-peer (P2P) approach, because millions of dedicated smartrouters are operated by the centralized video service providers in a coordinative manner. Thus it is intriguing to study the content placement strategies used in a smartrouter-based content delivery system, as well as its potential impact on the content delivery ecosystem. In this paper, we carry out measurement studies of Youku's peer video CDN, who has deployed over 300K smartrouter devices for its video delivery. In our measurement studies, 104K videos were investigated and 4TB traffic has been analyzed, over controlled smartrouter nodes and players. Our measurement insights are as follows. First, a global content replication strategy is essential for the peer CDN systems. Second, such peer CDN deployment itself can form an effective sub-system for end-to-end QoS monitoring, which can be used for fine-grained request redirection (e.g., user-level) and content replication. We also show our analysis on the performance limitations and propose potential improvements to the peer CDN systems.

Ming Ma, Zhi Wang, Ke Su et al.

Recent years have witnessed a new video delivery paradigm: smartrouter-based video delivery network, which is enabled by smartrouters deployed at users' homes, together with the conventional video servers deployed in the datacenters. Recently, ChinaCache, a large content delivery network (CDN) provider, and Youku, a video service provider using smartrouters to assist video delivery, announced their cooperation to create a new paradigm of content delivery based on householders' network resources. This new paradigm is different from the conventional peer-to-peer (P2P) approach, because such dedicated smartrouters are inherently operated by the centralized video service providers in a coordinative manner. It is intriguing to study the strategies, performance and potential impact on the content delivery ecosystem of such peer CDN systems. In this paper, we study the Youku peer CDN, which has deployed over 300K smartrouter devices for its video streaming. In our measurement, 78K videos were investigated and 3TB traffic has been analyzed, over controlled routers and players. Our contributions are the following measurement insights. First, a global replication and caching strategy is essential for the peer CDN systems, and proactively scheduling replication and caching on a daily basis can guarantee their performance. Second, such peer CDN deployment can itself form an effective Quality of Service (QoS) monitoring sub-system, which can be used for fine-grained user request redirection. We also provide our analysis on the performance issues and potential improvements to the peer CDN systems.

Zhi Wang

The Internet has recently witnessed the convergence of online social network services and online video services: users import videos from content sharing sites, and propagate them along the social connections by re-sharing them. Such social behaviors have dramatically reshaped how videos are disseminated, and the users are now actively engaged to be part of the social ecosystem, rather than being passively consumers. Despite the increasingly abundant bandwidth and computation resources, the ever increasing data volume of user generated video content and the boundless coverage of socialized sharing have presented unprecedented challenges. In this paper, we first presents the challenges in social-aware video delivery. Then, we present a principal framework for data-driven social video delivery approaches. Moreover, we identify the unique characteristics of social-aware video access and the social content propagation, and closely reveal the design of individual modules and their integration towards enhancing users' experience in the social network context.