Yang Yu

Ziniu Li, Tian Xu, Yushun Zhang et al.

Reinforcement Learning from Human Feedback (RLHF) is key to aligning Large Language Models (LLMs), typically paired with the Proximal Policy Optimization (PPO) algorithm. While PPO is a powerful method designed for general reinforcement learning tasks, it is overly sophisticated for LLMs, leading to laborious hyper-parameter tuning and significant computation burdens. To make RLHF efficient, we present ReMax, which leverages 3 properties of RLHF: fast simulation, deterministic transitions, and trajectory-level rewards. These properties are not exploited in PPO, making it less suitable for RLHF. Building on the renowned REINFORCE algorithm, ReMax does not require training an additional value model as in PPO and is further enhanced with a new variance reduction technique. ReMax offers several benefits over PPO: it is simpler to implement, eliminates more than 4 hyper-parameters in PPO, reduces GPU memory usage, and shortens training time. ReMax can save about 46% GPU memory than PPO when training a 7B model and enables training on A800-80GB GPUs without the memory-saving offloading technique needed by PPO. Applying ReMax to a Mistral-7B model resulted in a 94.78% win rate on the AlpacaEval leaderboard and a 7.739 score on MT-bench, setting a new SOTA for open-source 7B models. These results show the effectiveness of ReMax while addressing the limitations of PPO in LLMs.

Bo Wang, Jian Li, Yang Yu et al.

Considering that scene flow estimation has the capability of the spatial domain to focus but lacks the coherence of the temporal domain, this study proposes long-term scene flow estimation (LSFE), a comprehensive task that can simultaneously capture the fine-grained and long-term 3D motion in an online manner. We introduce SceneTracker, the first LSFE network that adopts an iterative approach to approximate the optimal 3D trajectory. The network dynamically and simultaneously indexes and constructs appearance correlation and depth residual features. Transformers are then employed to explore and utilize long-range connections within and between trajectories. With detailed experiments, SceneTracker shows superior capabilities in addressing 3D spatial occlusion and depth noise interference, highly tailored to the needs of the LSFE task. We build a real-world evaluation dataset, LSFDriving, for the LSFE field and use it in experiments to further demonstrate the advantage of SceneTracker in generalization abilities. The code and data are available at https://github.com/wwsource/SceneTracker.

Zhi-Hao Tan, Jian-Dong Liu, Xiao-Dong Bi et al.

The learnware paradigm proposed by Zhou [2016] aims to enable users to reuse numerous existing well-trained models instead of building machine learning models from scratch, with the hope of solving new user tasks even beyond models' original purposes. In this paradigm, developers worldwide can submit their high-performing models spontaneously to the learnware dock system (formerly known as learnware market) without revealing their training data. Once the dock system accepts the model, it assigns a specification and accommodates the model. This specification allows the model to be adequately identified and assembled to reuse according to future users' needs, even if they have no prior knowledge of the model. This paradigm greatly differs from the current big model direction and it is expected that a learnware dock system housing millions or more high-performing models could offer excellent capabilities for both planned tasks where big models are applicable; and unplanned, specialized, data-sensitive scenarios where big models are not present or applicable. This paper describes Beimingwu, the first open-source learnware dock system providing foundational support for future research of learnware paradigm.The system significantly streamlines the model development for new user tasks, thanks to its integrated architecture and engine design, extensive engineering implementations and optimizations, and the integration of various algorithms for learnware identification and reuse. Notably, this is possible even for users with limited data and minimal expertise in machine learning, without compromising the raw data's security. Beimingwu supports the entire process of learnware paradigm. The system lays the foundation for future research in learnware-related algorithms and systems, and prepares the ground for hosting a vast array of learnwares and establishing a learnware ecosystem.

Ruichu Cai, Siyang Huang, Jie Qiao et al.

As a key component to intuitive cognition and reasoning solutions in human intelligence, causal knowledge provides great potential for reinforcement learning (RL) agents' interpretability towards decision-making by helping reduce the searching space. However, there is still a considerable gap in discovering and incorporating causality into RL, which hinders the rapid development of causal RL. In this paper, we consider explicitly modeling the generation process of states with the causal graphical model, based on which we augment the policy. We formulate the causal structure updating into the RL interaction process with active intervention learning of the environment. To optimize the derived objective, we propose a framework with theoretical performance guarantees that alternates between two steps: using interventions for causal structure learning during exploration and using the learned causal structure for policy guidance during exploitation. Due to the lack of public benchmarks that allow direct intervention in the state space, we design the root cause localization task in our simulated fault alarm environment and then empirically show the effectiveness and robustness of the proposed method against state-of-the-art baselines. Theoretical analysis shows that our performance improvement attributes to the virtuous cycle of causal-guided policy learning and causal structure learning, which aligns with our experimental results. Codes are available at https://github.com/DMIRLAB-Group/FaultAlarm_RL.

Ziniu Li, Tian Xu, Yang Yu

Aligning intelligent agents with human preferences and values is important. This paper examines two popular alignment methods: Direct Preference Optimization (DPO) and Reward-Model-Based Policy Optimization (RMB-PO). A variant of RMB-PO, referred to as RMB-PO+ is also considered. These methods, either explicitly or implicitly, learn a reward model from preference data and differ in the data used for policy optimization to unlock the generalization ability of the reward model. In particular, compared with DPO, RMB-PO additionally uses policy-generated data, and RMB-PO+ further leverages new, preference-free data. We examine the impact of such out-of-preference data. Our study, conducted through controlled and synthetic experiments, demonstrates that DPO performs poorly, whereas RMB-PO+ performs the best. In particular, even when providing the policy model with a good feature representation, we find that policy optimization with adequate out-of-preference data significantly improves performance by harnessing the reward model's generalization capabilities.

Chengxing Jia, Fuxiang Zhang, Yi-Chen Li et al.

Offline meta-reinforcement learning (OMRL) proficiently allows an agent to tackle novel tasks while solely relying on a static dataset. For precise and efficient task identification, existing OMRL research suggests learning separate task representations that be incorporated with policy input, thus forming a context-based meta-policy. A major approach to train task representations is to adopt contrastive learning using multi-task offline data. The dataset typically encompasses interactions from various policies (i.e., the behavior policies), thus providing a plethora of contextual information regarding different tasks. Nonetheless, amassing data from a substantial number of policies is not only impractical but also often unattainable in realistic settings. Instead, we resort to a more constrained yet practical scenario, where multi-task data collection occurs with a limited number of policies. We observed that learned task representations from previous OMRL methods tend to correlate spuriously with the behavior policy instead of reflecting the essential characteristics of the task, resulting in unfavorable out-of-distribution generalization. To alleviate this issue, we introduce a novel algorithm to disentangle the impact of behavior policy from task representation learning through a process called adversarial data augmentation. Specifically, the objective of adversarial data augmentation is not merely to generate data analogous to offline data distribution; instead, it aims to create adversarial examples designed to confound learned task representations and lead to incorrect task identification. Our experiments show that learning from such adversarial samples significantly enhances the robustness and effectiveness of the task identification process and realizes satisfactory out-of-distribution generalization.

Peiyi Zhang, Xiaodong Jiang, Ginger M Holt et al.

Hyper-parameters of time series models play an important role in time series analysis. Slight differences in hyper-parameters might lead to very different forecast results for a given model, and therefore, selecting good hyper-parameter values is indispensable. Most of the existing generic hyper-parameter tuning methods, such as Grid Search, Random Search, Bayesian Optimal Search, are based on one key component - search, and thus they are computationally expensive and cannot be applied to fast and scalable time-series hyper-parameter tuning (HPT). We propose a self-supervised learning framework for HPT (SSL-HPT), which uses time series features as inputs and produces optimal hyper-parameters. SSL-HPT algorithm is 6-20x faster at getting hyper-parameters compared to other search based algorithms while producing comparable accurate forecasting results in various applications.

Yang Yu, Rongrong Ni, Yao Zhao

Recently, AI-manipulated face techniques have developed rapidly and constantly, which has raised new security issues in society. Although existing detection methods consider different categories of fake faces, the performance on detecting the fake faces with "unseen" manipulation techniques is still poor due to the distribution bias among cross-manipulation techniques. To solve this problem, we propose a novel framework that focuses on mining intrinsic features and further eliminating the distribution bias to improve the generalization ability. Firstly, we focus on mining the intrinsic clues in the channel difference image (CDI) and spectrum image (SI) from the camera imaging process and the indispensable step in AI manipulation process. Then, we introduce the Octave Convolution (OctConv) and an attention-based fusion module to effectively and adaptively mine intrinsic features from CDI and SI. Finally, we design an alignment module to eliminate the bias of manipulation techniques to obtain a more generalized detection framework. We evaluate the proposed framework on four categories of fake faces datasets with the most popular and state-of-the-art manipulation techniques, and achieve very competitive performances. To further verify the generalization ability of the proposed framework, we conduct experiments on cross-manipulation techniques, and the results show the advantages of our method.

Zenan Ling, Haotian Ma, Yu Yang et al.

In this paper, we propose to disentangle and interpret contextual effects that are encoded in a pre-trained deep neural network. We use our method to explain the gaming strategy of the alphaGo Zero model. Unlike previous studies that visualized image appearances corresponding to the network output or a neural activation only from a global perspective, our research aims to clarify how a certain input unit (dimension) collaborates with other units (dimensions) to constitute inference patterns of the neural network and thus contribute to the network output. The analysis of local contextual effects w.r.t. certain input units is of special values in real applications. Explaining the logic of the alphaGo Zero model is a typical application. In experiments, our method successfully disentangled the rationale of each move during the Go game.