Bo Li

Zhaorun Chen, Zhuokai Zhao, Kai Zhang et al.

While reinforcement learning (RL) can empower autonomous agents by enabling self-improvement through interaction, its practical adoption remains challenging due to costly rollouts, limited task diversity, unreliable reward signals, and infrastructure complexity, all of which obstruct the collection of scalable experience data. To address these challenges, we introduce DreamGym, the first unified framework designed to synthesize diverse experiences with scalability in mind to enable effective online RL training for autonomous agents. Rather than relying on expensive real-environment rollouts, DreamGym distills environment dynamics into a reasoning-based experience model that derives consistent state transitions and feedback signals through step-by-step reasoning, enabling scalable agent rollout collection for RL. To improve the stability and quality of transitions, DreamGym leverages an experience replay buffer initialized with offline real-world data and continuously enriched with fresh interactions to actively support agent training. To improve knowledge acquisition, DreamGym adaptively generates new tasks that challenge the current agent policy, enabling more effective online curriculum learning. Experiments across diverse environments and agent backbones demonstrate that DreamGym substantially improves RL training, both in fully synthetic settings and in sim-to-real transfer scenarios. On non-RL-ready tasks like WebArena, DreamGym outperforms all baselines by over 30%. And in RL-ready but costly settings, it matches GRPO and PPO performance using only synthetic interactions. When transferring a policy trained purely on synthetic experiences to real-environment RL, DreamGym yields significant additional performance gains while requiring far fewer real-world interactions, providing a scalable warm-start strategy for general-purpose RL.

Yulong Fan, Rong Chen, Bo Li

Rain removal in images is an important task in computer vision filed and attracting attentions of more and more people. In this paper, we address a non-trivial issue of removing visual effect of rain streak from a single image. Differing from existing work, our method combines various semantic constraint task in a proposed multi-task regression model for rain removal. These tasks reinforce the model's capabilities from the content, edge-aware, and local texture similarity respectively. To further improve the performance of multi-task learning, we also present two simple but powerful dynamic weighting algorithms. The proposed multi-task enhanced network (MENET) is a powerful convolutional neural network based on U-Net for rain removal research, with a specific focus on utilize multiple tasks constraints and exploit the synergy among them to facilitate the model's rain removal capacity. It is noteworthy that the adaptive weighting scheme has further resulted in improved network capability. We conduct several experiments on synthetic and real rain images, and achieve superior rain removal performance over several selected state-of-the-art (SOTA) approaches. The overall effect of our method is impressive, even in the decomposition of heavy rain and rain streak accumulation.The source code and some results can be found at:https://github.com/SumiHui/MENET.

Xinglin Pan, Wenxiang Lin, Lin Zhang et al.

Recent large language models (LLMs) have tended to leverage sparsity to reduce computations, employing the sparsely activated mixture-of-experts (MoE) technique. MoE introduces four modules, including token routing, token communication, expert computation, and expert parallelism, that impact model quality and training efficiency. To enable versatile usage of MoE models, we introduce FSMoE, a flexible training system optimizing task scheduling with three novel techniques: 1) Unified abstraction and online profiling of MoE modules for task scheduling across various MoE implementations. 2) Co-scheduling intra-node and inter-node communications with computations to minimize communication overheads. 3) To support near-optimal task scheduling, we design an adaptive gradient partitioning method for gradient aggregation and a schedule to adaptively pipeline communications and computations. We conduct extensive experiments with configured MoE layers and real-world MoE models on two GPU clusters. Experimental results show that 1) our FSMoE supports four popular types of MoE routing functions and is more efficient than existing implementations (with up to a 1.42$\times$ speedup), and 2) FSMoE outperforms the state-of-the-art MoE training systems (DeepSpeed-MoE and Tutel) by 1.18$\times$-1.22$\times$ on 1458 MoE layers and 1.19$\times$-3.01$\times$ on real-world MoE models based on GPT-2 and Mixtral using a popular routing function.

Liu Yuezhang, Bo Li, Qifeng Chen

It is well known that artificial neural networks are vulnerable to adversarial examples, in which great efforts have been made to improve the robustness. However, such examples are usually imperceptible to humans, and thus their effect on biological neural circuits is largely unknown. This paper will investigate the adversarial robustness in a simulated cerebellum, a well-studied supervised learning system in computational neuroscience. Specifically, we propose to study three unique characteristics revealed in the cerebellum: (i) network width; (ii) long-term depression on the parallel fiber-Purkinje cell synapses; (iii) sparse connectivity in the granule layer, and hypothesize that they will be beneficial for improving robustness. To the best of our knowledge, this is the first attempt to examine the adversarial robustness in simulated cerebellum models. The results are negative in the experimental phase -- no significant improvements in robustness are discovered from the proposed three mechanisms. Consequently, the cerebellum is expected to be vulnerable to adversarial examples as the deep neural networks under batch training. Neuroscientists are encouraged to fool the biological system in experiments with adversarial attacks.

Shaohuai Shi, Zhenheng Tang, Xiaowen Chu et al.

Nowadays, large and complex deep learning (DL) models are increasingly trained in a distributed manner across multiple worker machines, in which extensive communications between workers pose serious scaling problems. In this article, we present a quantitative survey of communication optimization techniques for data parallel distributed DL. We first identify the major communication challenges and classify the existing solutions into three levels, namely the learning algorithm, the system architecture, and the network infrastructure. We present the state-of-the-art communication optimization techniques and conduct a comparative study of seven common lossless distributed DL methods on a 32-GPU cluster with 100Gbps InfiniBand (IB). We show that (1) the DL models with low model intensity (such as BERT and BERT-Large) are difficult to scale out even with the best available lossless algorithm over 100Gbps IB; (2) the system architecture and scheduling algorithms have a critical impact on the scaling property. We conclude the article with discussions on the open issues for further investigations.

Tianyang Wang, Jun Huan, Bo Li

Deep learning models learn to fit training data while they are highly expected to generalize well to testing data. Most works aim at finding such models by creatively designing architectures and fine-tuning parameters. To adapt to particular tasks, hand-crafted information such as image prior has also been incorporated into end-to-end learning. However, very little progress has been made on investigating how an individual training sample will influence the generalization ability of a model. In other words, to achieve high generalization accuracy, do we really need all the samples in a training dataset? In this paper, we demonstrate that deep learning models such as convolutional neural networks may not favor all training samples, and generalization accuracy can be further improved by dropping those unfavorable samples. Specifically, the influence of removing a training sample is quantifiable, and we propose a Two-Round Training approach, aiming to achieve higher generalization accuracy. We locate unfavorable samples after the first round of training, and then retrain the model from scratch with the reduced training dataset in the second round. Since our approach is essentially different from fine-tuning or further training, the computational cost should not be a concern. Our extensive experimental results indicate that, with identical settings, the proposed approach can boost performance of the well-known networks on both high-level computer vision problems such as image classification, and low-level vision problems such as image denoising.

Lei Huang, Xianglong Liu, Bo Lang et al.

Optimizing deep neural networks (DNNs) often suffers from the ill-conditioned problem. We observe that the scaling-based weight space symmetry property in rectified nonlinear network will cause this negative effect. Therefore, we propose to constrain the incoming weights of each neuron to be unit-norm, which is formulated as an optimization problem over Oblique manifold. A simple yet efficient method referred to as projection based weight normalization (PBWN) is also developed to solve this problem. PBWN executes standard gradient updates, followed by projecting the updated weight back to Oblique manifold. This proposed method has the property of regularization and collaborates well with the commonly used batch normalization technique. We conduct comprehensive experiments on several widely-used image datasets including CIFAR-10, CIFAR-100, SVHN and ImageNet for supervised learning over the state-of-the-art convolutional neural networks, such as Inception, VGG and residual networks. The results show that our method is able to improve the performance of DNNs with different architectures consistently. We also apply our method to Ladder network for semi-supervised learning on permutation invariant MNIST dataset, and our method outperforms the state-of-the-art methods: we obtain test errors as 2.52%, 1.06%, and 0.91% with only 20, 50, and 100 labeled samples, respectively.

Bo Li, Huahui Chen, Yucheng Chen et al.

Action recognition from well-segmented 3D skeleton video has been intensively studied. However, due to the difficulty in representing the 3D skeleton video and the lack of training data, action detection from streaming 3D skeleton video still lags far behind its recognition counterpart and image based object detection. In this paper, we propose a novel approach for this problem, which leverages both effective skeleton video encoding and deep regression based object detection from images. Our framework consists of two parts: skeleton-based video image mapping, which encodes a skeleton video to a color image in a temporal preserving way, and an end-to-end trainable fast skeleton action detector (Skeleton Boxes) based on image detection. Experimental results on the latest and largest PKU-MMD benchmark dataset demonstrate that our method outperforms the state-of-the-art methods with a large margin. We believe our idea would inspire and benefit future research in this important area.