Haichao Zhang

Lingfeng Sun, Haichao Zhang, Wei Xu et al. · berkeley

The purpose of multi-task reinforcement learning (MTRL) is to train a single policy that can be applied to a set of different tasks. Sharing parameters allows us to take advantage of the similarities among tasks. However, the gaps between contents and difficulties of different tasks bring us challenges on both which tasks should share the parameters and what parameters should be shared, as well as the optimization challenges due to parameter sharing. In this work, we introduce a parameter-compositional approach (PaCo) as an attempt to address these challenges. In this framework, a policy subspace represented by a set of parameters is learned. Policies for all the single tasks lie in this subspace and can be composed by interpolating with the learned set. It allows not only flexible parameter sharing but also a natural way to improve training. We demonstrate the state-of-the-art performance on Meta-World benchmarks, verifying the effectiveness of the proposed approach.

Lingfeng Sun, Haichao Zhang, Wei Xu et al. · berkeley

In this work, we investigate the potential of improving multi-task training and also leveraging it for transferring in the reinforcement learning setting. We identify several challenges towards this goal and propose a transferring approach with a parameter-compositional formulation. We investigate ways to improve the training of multi-task reinforcement learning which serves as the foundation for transferring. Then we conduct a number of transferring experiments on various manipulation tasks. Experimental results demonstrate that the proposed approach can have improved performance in the multi-task training stage, and further show effective transferring in terms of both sample efficiency and performance.

Haichao Zhang, We Xu, Haonan Yu

Pre-training with offline data and online fine-tuning using reinforcement learning is a promising strategy for learning control policies by leveraging the best of both worlds in terms of sample efficiency and performance. One natural approach is to initialize the policy for online learning with the one trained offline. In this work, we introduce a policy expansion scheme for this task. After learning the offline policy, we use it as one candidate policy in a policy set. We then expand the policy set with another policy which will be responsible for further learning. The two policies will be composed in an adaptive manner for interacting with the environment. With this approach, the policy previously learned offline is fully retained during online learning, thus mitigating the potential issues such as destroying the useful behaviors of the offline policy in the initial stage of online learning while allowing the offline policy participate in the exploration naturally in an adaptive manner. Moreover, new useful behaviors can potentially be captured by the newly added policy through learning. Experiments are conducted on a number of tasks and the results demonstrate the effectiveness of the proposed approach. Code is available at https://github.com/Haichao-Zhang/PEX

Haichao Zhang, Jiashi Li, Xin Xia et al.

Recently, Neural architecture search has achieved great success on classification tasks for mobile devices. The backbone network for object detection is usually obtained on the image classification task. However, the architecture which is searched through the classification task is sub-optimal because of the gap between the task of image and object detection. As while work focuses on backbone network architecture search for mobile device object detection is limited, mainly because the backbone always requires expensive ImageNet pre-training. Accordingly, it is necessary to study the approach of network architecture search for mobile device object detection without expensive pre-training. In this work, we propose a mobile object detection backbone network architecture search algorithm which is a kind of evolutionary optimized method based on non-dominated sorting for NAS scenarios. It can quickly search to obtain the backbone network architecture within certain constraints. It better solves the problem of suboptimal linear combination accuracy and computational cost. The proposed approach can search the backbone networks with different depths, widths, or expansion sizes via a technique of weight mapping, making it possible to use NAS for mobile devices detection tasks a lot more efficiently. In our experiments, we verify the effectiveness of the proposed approach on YoloX-Lite, a lightweight version of the target detection framework. Under similar computational complexity, the accuracy of the backbone network architecture we search for is 2.0% mAP higher than MobileDet. Our improved backbone network can reduce the computational effort while improving the accuracy of the object detection network. To prove its effectiveness, a series of ablation studies have been carried out and the working mechanism has been analyzed in detail.

Haichao Zhang, Can Qin, Yu Yin et al.

Camouflaged objects that blend into natural scenes pose significant challenges for deep-learning models to detect and synthesize. While camouflaged object detection is a crucial task in computer vision with diverse real-world applications, this research topic has been constrained by limited data availability. We propose a framework for synthesizing camouflage data to enhance the detection of camouflaged objects in natural scenes. Our approach employs a generative model to produce realistic camouflage images, which can be used to train existing object detection models. Specifically, we use a camouflage environment generator supervised by a camouflage distribution classifier to synthesize the camouflage images, which are then fed into our generator to expand the dataset. Our framework outperforms the current state-of-the-art method on three datasets (COD10k, CAMO, and CHAMELEON), demonstrating its effectiveness in improving camouflaged object detection. This approach can serve as a plug-and-play data generation and augmentation module for existing camouflaged object detection tasks and provides a novel way to introduce more diversity and distributions into current camouflage datasets.

Yuwei Fu, Haichao Zhang, Di Wu et al.

Reward specification is one of the most tricky problems in Reinforcement Learning, which usually requires tedious hand engineering in practice. One promising approach to tackle this challenge is to adopt existing expert video demonstrations for policy learning. Some recent work investigates how to learn robot policies from only a single/few expert video demonstrations. For example, reward labeling via Optimal Transport (OT) has been shown to be an effective strategy to generate a proxy reward by measuring the alignment between the robot trajectory and the expert demonstrations. However, previous work mostly overlooks that the OT reward is invariant to temporal order information, which could bring extra noise to the reward signal. To address this issue, in this paper, we introduce the Temporal Optimal Transport (TemporalOT) reward to incorporate temporal order information for learning a more accurate OT-based proxy reward. Extensive experiments on the Meta-world benchmark tasks validate the efficacy of the proposed method. Code is available at: https://github.com/fuyw/TemporalOT

Haichao Zhang, Chong Zhang, Peiyu Hu et al.

Modern recommender systems face a critical challenge in complying with privacy regulations like the 'right to be forgotten': removing a user's data without disrupting recommendations for others. Traditional unlearning methods address this by partial model updates, but introduce propagation bias--where unlearning one user's data distorts recommendations for behaviorally similar users, degrading system accuracy. While retraining eliminates bias, it is computationally prohibitive for large-scale systems. To address this challenge, we propose CRAGRU, a novel framework leveraging Retrieval-Augmented Generation (RAG) for efficient, user-specific unlearning that mitigates bias while preserving recommendation quality. CRAGRU decouples unlearning into distinct retrieval and generation stages. In retrieval, we employ three tailored strategies designed to precisely isolate the target user's data influence, minimizing collateral impact on unrelated users and enhancing unlearning efficiency. Subsequently, the generation stage utilizes an LLM, augmented with user profiles integrated into prompts, to reconstruct accurate and personalized recommendations without needing to retrain the entire base model. Experiments on three public datasets demonstrate that CRAGRU effectively unlearns targeted user data, significantly mitigating unlearning bias by preventing adverse impacts on non-target users, while maintaining recommendation performance comparable to fully trained original models. Our work highlights the promise of RAG-based architectures for building robust and privacy-preserving recommender systems. The source code is available at: https://github.com/zhanghaichao520/LLM_rec_unlearning.

Yuwei Fu, Haichao Zhang, Di Wu et al.

In this work, we investigate how to leverage pre-trained visual-language models (VLM) for online Reinforcement Learning (RL). In particular, we focus on sparse reward tasks with pre-defined textual task descriptions. We first identify the problem of reward misalignment when applying VLM as a reward in RL tasks. To address this issue, we introduce a lightweight fine-tuning method, named Fuzzy VLM reward-aided RL (FuRL), based on reward alignment and relay RL. Specifically, we enhance the performance of SAC/DrQ baseline agents on sparse reward tasks by fine-tuning VLM representations and using relay RL to avoid local minima. Extensive experiments on the Meta-world benchmark tasks demonstrate the efficacy of the proposed method. Code is available at: https://github.com/fuyw/FuRL.

Haonan Yu, Wei Xu, Haichao Zhang

We consider the safe reinforcement learning (RL) problem of maximizing utility with extremely low constraint violation rates. Assuming no prior knowledge or pre-training of the environment safety model given a task, an agent has to learn, via exploration, which states and actions are safe. A popular approach in this line of research is to combine a model-free RL algorithm with the Lagrangian method to adjust the weight of the constraint reward relative to the utility reward dynamically. It relies on a single policy to handle the conflict between utility and constraint rewards, which is often challenging. We present SEditor, a two-policy approach that learns a safety editor policy transforming potentially unsafe actions proposed by a utility maximizer policy into safe ones. The safety editor is trained to maximize the constraint reward while minimizing a hinge loss of the utility state-action values before and after an action is edited. SEditor extends existing safety layer designs that assume simplified safety models, to general safe RL scenarios where the safety model can in theory be arbitrarily complex. As a first-order method, it is easy to implement and efficient for both inference and training. On 12 Safety Gym tasks and 2 safe racing tasks, SEditor obtains much a higher overall safety-weighted-utility (SWU) score than the baselines, and demonstrates outstanding utility performance with constraint violation rates as low as once per 2k time steps, even in obstacle-dense environments. On some tasks, this low violation rate is up to 200 times lower than that of an unconstrained RL method with similar utility performance. Code is available at https://github.com/hnyu/seditor.

Haonan Yu, Wei Xu, Haichao Zhang

We present temporally abstract actor-critic (TAAC), a simple but effective off-policy RL algorithm that incorporates closed-loop temporal abstraction into the actor-critic framework. TAAC adds a second-stage binary policy to choose between the previous action and a new action output by an actor. Crucially, its "act-or-repeat" decision hinges on the actually sampled action instead of the expected behavior of the actor. This post-acting switching scheme let the overall policy make more informed decisions. TAAC has two important features: a) persistent exploration, and b) a new compare-through Q operator for multi-step TD backup, specially tailored to the action repetition scenario. We demonstrate TAAC's advantages over several strong baselines across 14 continuous control tasks. Our surprising finding reveals that while achieving top performance, TAAC is able to "mine" a significant number of repeated actions with the trained policy even on continuous tasks whose problem structures on the surface seem to repel action repetition. This suggests that aside from encouraging persistent exploration, action repetition can find its place in a good policy behavior. Code is available at https://github.com/hnyu/taac.

Zhuo Li, Xianghuai Deng, Chiwei Feng et al.

Large language models (LLMs) have significantly reshaped different walks of business. To meet the increasing demands for individualized railway service, we develop LLM4Rail - a novel LLM-augmented railway service consulting platform. Empowered by LLM, LLM4Rail can provide custom modules for ticketing, railway food & drink recommendations, weather information, and chitchat. In LLM4Rail, we propose the iterative "Question-Thought-Action-Observation (QTAO)" prompting framework. It meticulously integrates verbal reasoning with task-oriented actions, that is, reasoning to guide action selection, to effectively retrieve external observations relevant to railway operation and service to generate accurate responses. To provide personalized onboard dining services, we first construct the Chinese Railway Food and Drink (CRFD-25) - a publicly accessible takeout dataset tailored for railway services. CRFD-25 covers a wide range of signature dishes categorized by cities, cuisines, age groups, and spiciness levels. We further introduce an LLM-based zero-shot conversational recommender for railway catering. To address the unconstrained nature of open recommendations, the feature similarity-based post-processing step is introduced to ensure all the recommended items are aligned with CRFD-25 dataset.

Haonan Yu, Haichao Zhang, Wei Xu

Maximum entropy (MaxEnt) RL maximizes a combination of the original task reward and an entropy reward. It is believed that the regularization imposed by entropy, on both policy improvement and policy evaluation, together contributes to good exploration, training convergence, and robustness of learned policies. This paper takes a closer look at entropy as an intrinsic reward, by conducting various ablation studies on soft actor-critic (SAC), a popular representative of MaxEnt RL. Our findings reveal that in general, entropy rewards should be applied with caution to policy evaluation. On one hand, the entropy reward, like any other intrinsic reward, could obscure the main task reward if it is not properly managed. We identify some failure cases of the entropy reward especially in episodic Markov decision processes (MDPs), where it could cause the policy to be overly optimistic or pessimistic. On the other hand, our large-scale empirical study shows that using entropy regularization alone in policy improvement, leads to comparable or even better performance and robustness than using it in both policy improvement and policy evaluation. Based on these observations, we recommend either normalizing the entropy reward to a zero mean (SACZero), or simply removing it from policy evaluation (SACLite) for better practical results.

Haichao Zhang, Wei Xu, Haonan Yu

Standard model-free reinforcement learning algorithms optimize a policy that generates the action to be taken in the current time step in order to maximize expected future return. While flexible, it faces difficulties arising from the inefficient exploration due to its single step nature. In this work, we present Generative Planning method (GPM), which can generate actions not only for the current step, but also for a number of future steps (thus termed as generative planning). This brings several benefits to GPM. Firstly, since GPM is trained by maximizing value, the plans generated from it can be regarded as intentional action sequences for reaching high value regions. GPM can therefore leverage its generated multi-step plans for temporally coordinated exploration towards high value regions, which is potentially more effective than a sequence of actions generated by perturbing each action at single step level, whose consistent movement decays exponentially with the number of exploration steps. Secondly, starting from a crude initial plan generator, GPM can refine it to be adaptive to the task, which, in return, benefits future explorations. This is potentially more effective than commonly used action-repeat strategy, which is non-adaptive in its form of plans. Additionally, since the multi-step plan can be interpreted as the intent of the agent from now to a span of time period into the future, it offers a more informative and intuitive signal for interpretation. Experiments are conducted on several benchmark environments and the results demonstrated its effectiveness compared with several baseline methods.

Chenxing Wang, Fang Zhao, Haichao Zhang et al.

Travel Time Estimation (TTE) is indispensable in intelligent transportation system (ITS). It is significant to achieve the fine-grained Trajectory-based Travel Time Estimation (TTTE) for multi-city scenarios, namely to accurately estimate travel time of the given trajectory for multiple city scenarios. However, it faces great challenges due to complex factors including dynamic temporal dependencies and fine-grained spatial dependencies. To tackle these challenges, we propose a meta learning based framework, MetaTTE, to continuously provide accurate travel time estimation over time by leveraging well-designed deep neural network model called DED, which consists of Data preprocessing module and Encoder-Decoder network module. By introducing meta learning techniques, the generalization ability of MetaTTE is enhanced using small amount of examples, which opens up new opportunities to increase the potential of achieving consistent performance on TTTE when traffic conditions and road networks change over time in the future. The DED model adopts an encoder-decoder network to capture fine-grained spatial and temporal representations. Extensive experiments on two real-world datasets are conducted to confirm that our MetaTTE outperforms six state-of-art baselines, and improve 29.35% and 25.93% accuracy than the best baseline on Chengdu and Porto datasets, respectively.

Wei Wei, Haonan Yu, Haichao Zhang et al.

In robot sensing scenarios, instead of passively utilizing human captured views, an agent should be able to actively choose informative viewpoints of a 3D object as discriminative evidence to boost the recognition accuracy. This task is referred to as active object recognition. Recent works on this task rely on a massive amount of training examples to train an optimal view selection policy. But in realistic robot sensing scenarios, the large-scale training data may not exist and whether the intelligent view selection policy can be still learned from few object samples remains unclear. In this paper, we study this new problem which is extremely challenging but very meaningful in robot sensing -- Few-shot Active Object Recognition, i.e., to learn view selection policies from few object samples, which has not been considered and addressed before. We solve the proposed problem by adopting the framework of meta learning and name our method "MetaView". Extensive experiments on both category-level and instance-level classification tasks demonstrate that the proposed method can efficiently resolve issues that are hard for state-of-the-art active object recognition methods to handle, and outperform several baselines by large margins.

Haichao Zhang, Jianyu Wang

We introduce a feature scattering-based adversarial training approach for improving model robustness against adversarial attacks. Conventional adversarial training approaches leverage a supervised scheme (either targeted or non-targeted) in generating attacks for training, which typically suffer from issues such as label leaking as noted in recent works. Differently, the proposed approach generates adversarial images for training through feature scattering in the latent space, which is unsupervised in nature and avoids label leaking. More importantly, this new approach generates perturbed images in a collaborative fashion, taking the inter-sample relationships into consideration. We conduct analysis on model robustness and demonstrate the effectiveness of the proposed approach through extensively experiments on different datasets compared with state-of-the-art approaches.

Haichao Zhang, Jianyu Wang

Conventional adversarial training methods using attacks that manipulate the pixel value directly and individually, leading to models that are less robust in face of spatial transformation-based attacks. In this paper, we propose a joint adversarial training method that incorporates both spatial transformation-based and pixel-value based attacks for improving model robustness. We introduce a spatial transformation-based attack with an explicit notion of budget and develop an algorithm for spatial attack generation. We further integrate both pixel and spatial attacks into one generation model and show how to leverage the complementary strengths of each other in training for improving the overall model robustness. Extensive experimental results on different benchmark datasets compared with state-of-the-art methods verified the effectiveness of the proposed method.

Haichao Zhang, Jianyu Wang

Object detection is an important vision task and has emerged as an indispensable component in many vision system, rendering its robustness as an increasingly important performance factor for practical applications. While object detection models have been demonstrated to be vulnerable against adversarial attacks by many recent works, very few efforts have been devoted to improving their robustness. In this work, we take an initial attempt towards this direction. We first revisit and systematically analyze object detectors and many recently developed attacks from the perspective of model robustness. We then present a multi-task learning perspective of object detection and identify an asymmetric role of task losses. We further develop an adversarial training approach which can leverage the multiple sources of attacks for improving the robustness of detection models. Extensive experiments on PASCAL-VOC and MS-COCO verified the effectiveness of the proposed approach.

Liqun Chen, Yizhe Zhang, Ruiyi Zhang et al.

Sequence-to-sequence models are commonly trained via maximum likelihood estimation (MLE). However, standard MLE training considers a word-level objective, predicting the next word given the previous ground-truth partial sentence. This procedure focuses on modeling local syntactic patterns, and may fail to capture long-range semantic structure. We present a novel solution to alleviate these issues. Our approach imposes global sequence-level guidance via new supervision based on optimal transport, enabling the overall characterization and preservation of semantic features. We further show that this method can be understood as a Wasserstein gradient flow trying to match our model to the ground truth sequence distribution. Extensive experiments are conducted to validate the utility of the proposed approach, showing consistent improvements over a wide variety of NLP tasks, including machine translation, abstractive text summarization, and image captioning.

Jianyu Wang, Haichao Zhang

In this paper, we study fast training of adversarially robust models. From the analyses of the state-of-the-art defense method, i.e., the multi-step adversarial training, we hypothesize that the gradient magnitude links to the model robustness. Motivated by this, we propose to perturb both the image and the label during training, which we call Bilateral Adversarial Training (BAT). To generate the adversarial label, we derive an closed-form heuristic solution. To generate the adversarial image, we use one-step targeted attack with the target label being the most confusing class. In the experiment, we first show that random start and the most confusing target attack effectively prevent the label leaking and gradient masking problem. Then coupled with the adversarial label part, our model significantly improves the state-of-the-art results. For example, against PGD100 white-box attack with cross-entropy loss, on CIFAR10, we achieve 63.7\% versus 47.2\%; on SVHN, we achieve 59.1\% versus 42.1\%. At last, the experiment on the very (computationally) challenging ImageNet dataset further demonstrates the effectiveness of our fast method.

Haichao Zhang, Haonan Yu, Wei Xu

Building intelligent agents that can communicate with and learn from humans in natural language is of great value. Supervised language learning is limited by the ability of capturing mainly the statistics of training data, and is hardly adaptive to new scenarios or flexible for acquiring new knowledge without inefficient retraining or catastrophic forgetting. We highlight the perspective that conversational interaction serves as a natural interface both for language learning and for novel knowledge acquisition and propose a joint imitation and reinforcement approach for grounded language learning through an interactive conversational game. The agent trained with this approach is able to actively acquire information by asking questions about novel objects and use the just-learned knowledge in subsequent conversations in a one-shot fashion. Results compared with other methods verified the effectiveness of the proposed approach.

Haichao Zhang, Haonan Yu, Wei Xu

One of the long-term goals of artificial intelligence is to build an agent that can communicate intelligently with human in natural language. Most existing work on natural language learning relies heavily on training over a pre-collected dataset with annotated labels, leading to an agent that essentially captures the statistics of the fixed external training data. As the training data is essentially a static snapshot representation of the knowledge from the annotator, the agent trained this way is limited in adaptiveness and generalization of its behavior. Moreover, this is very different from the language learning process of humans, where language is acquired during communication by taking speaking action and learning from the consequences of speaking action in an interactive manner. This paper presents an interactive setting for grounded natural language learning, where an agent learns natural language by interacting with a teacher and learning from feedback, thus learning and improving language skills while taking part in the conversation. To achieve this goal, we propose a model which incorporates both imitation and reinforcement by leveraging jointly sentence and reward feedbacks from the teacher. Experiments are conducted to validate the effectiveness of the proposed approach.

Haichao Zhang, David Wipf

Typical blur from camera shake often deviates from the standard uniform convolutional script, in part because of problematic rotations which create greater blurring away from some unknown center point. Consequently, successful blind deconvolution requires the estimation of a spatially-varying or non-uniform blur operator. Using ideas from Bayesian inference and convex analysis, this paper derives a non-uniform blind deblurring algorithm with several desirable, yet previously-unexplored attributes. The underlying objective function includes a spatially adaptive penalty which couples the latent sharp image, non-uniform blur operator, and noise level together. This coupling allows the penalty to automatically adjust its shape based on the estimated degree of local blur and image structure such that regions with large blur or few prominent edges are discounted. Remaining regions with modest blur and revealing edges therefore dominate the overall estimation process without explicitly incorporating structure-selection heuristics. The algorithm can be implemented using a majorization-minimization strategy that is virtually parameter free. Detailed theoretical analysis and empirical validation on real images serve to validate the proposed method.

David Wipf, Haichao Zhang

Blind deconvolution involves the estimation of a sharp signal or image given only a blurry observation. Because this problem is fundamentally ill-posed, strong priors on both the sharp image and blur kernel are required to regularize the solution space. While this naturally leads to a standard MAP estimation framework, performance is compromised by unknown trade-off parameter settings, optimization heuristics, and convergence issues stemming from non-convexity and/or poor prior selections. To mitigate some of these problems, a number of authors have recently proposed substituting a variational Bayesian (VB) strategy that marginalizes over the high-dimensional image space leading to better estimates of the blur kernel. However, the underlying cost function now involves both integrals with no closed-form solution and complex, function-valued arguments, thus losing the transparency of MAP. Beyond standard Bayesian-inspired intuitions, it thus remains unclear by exactly what mechanism these methods are able to operate, rendering understanding, improvements and extensions more difficult. To elucidate these issues, we demonstrate that the VB methodology can be recast as an unconventional MAP problem with a very particular penalty/prior that couples the image, blur kernel, and noise level in a principled way. This unique penalty has a number of useful characteristics pertaining to relative concavity, local minima avoidance, and scale-invariance that allow us to rigorously explain the success of VB including its existing implementational heuristics and approximations. It also provides strict criteria for choosing the optimal image prior that, perhaps counter-intuitively, need not reflect the statistics of natural scenes. In so doing we challenge the prevailing notion of why VB is successful for blind deconvolution while providing a transparent platform for introducing enhancements.

Haichao Zhang, David Wipf, Yanning Zhang

Image super-resolution (SR) is one of the long-standing and active topics in image processing community. A large body of works for image super resolution formulate the problem with Bayesian modeling techniques and then obtain its Maximum-A-Posteriori (MAP) solution, which actually boils down to a regularized regression task over separable regularization term. Although straightforward, this approach cannot exploit the full potential offered by the probabilistic modeling, as only the posterior mode is sought. Also, the separable property of the regularization term can not capture any correlations between the sparse coefficients, which sacrifices much on its modeling accuracy. We propose a Bayesian image SR algorithm via sparse modeling of natural images. The sparsity property of the latent high resolution image is exploited by introducing latent variables into the high-order Markov Random Field (MRF) which capture the content adaptive variance by pixel-wise adaptation. The high-resolution image is estimated via Empirical Bayesian estimation scheme, which is substantially faster than our previous approach based on Markov Chain Monte Carlo sampling [1]. It is shown that the actual cost function for the proposed approach actually incorporates a non-factorial regularization term over the sparse coefficients. Experimental results indicate that the proposed method can generate competitive or better results than \emph{state-of-the-art} SR algorithms.