Searcharxiv⌕ Search

arXiv subjects

Haoxiang Wang

Publications and source records attributed to Haoxiang Wang.

At least 37 records · Page 2Linked to original sources
arXiv2024

RLHF Workflow: From Reward Modeling to Online RLHF

We present the workflow of Online Iterative Reinforcement Learning from Human Feedback (RLHF) in this technical report, which is widely reported to outperform its offline counterpart by a large margin in the recent large language model (LLM) literature. However, existing open-source RLHF projects are still largely confined to the offline learning setting. In this technical report, we aim to fill in this gap and provide a detailed recipe that is easy to reproduce for online iterative RLHF. In particular, since online human feedback is usually infeasible for open-source communities with limited resources, we start by constructing preference models using a diverse set of open-source datasets and use the constructed proxy preference model to approximate human feedback. Then, we discuss the theoretical insights and algorithmic principles behind online iterative RLHF, followed by a detailed practical implementation. Our trained LLM achieves impressive performance on LLM chatbot benchmarks, including AlpacaEval-2, Arena-Hard, and MT-Bench, as well as other academic benchmarks such as HumanEval and TruthfulQA. We have shown that supervised fine-tuning (SFT) and iterative RLHF can obtain state-of-the-art performance with fully open-source datasets. Further, we have made our models, curated datasets, and comprehensive step-by-step code guidebooks publicly available. Please refer to https://github.com/RLHFlow/RLHF-Reward-Modeling and https://github.com/RLHFlow/Online-RLHF for more detailed information.

cs.LG↗
arXiv2024

Mitigating the Alignment Tax of RLHF

LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax. To investigate alignment tax, we conducted experiments with existing RLHF algorithms using OpenLLaMA-3B, which revealed a pronounced alignment tax in NLP tasks. Whereas, despite various techniques to mitigate forgetting, they are often at odds with the RLHF performance, leading to a trade-off between alignment performance and forgetting mitigation, leading to an alignment-forgetting trade-off. In this paper we show that model averaging, which simply interpolates between pre and post RLHF model weights, surprisingly achieves the most strongest alignment-forgetting Pareto front among a wide range of competing methods. To understand its effectiveness, we offer theoretical insights into model averaging, revealing that it enhances performance Pareto front by increasing feature diversity on the layers where tasks share overlapped feature spaces. Empirical evidence corroborates our analysis by showing the benefits of averaging low-level transformer layers. Building on the analysis and the observation that averaging different layers of the transformer leads to significantly different alignment-forgetting trade-offs, we propose Heterogeneous Model Averaging (HMA) to Heterogeneously find various combination ratios of model layers. HMA seeks to maximize the alignment performance while incurring minimal alignment tax. Moreover, we validate HMA's performance across a range of RLHF algorithms over OpenLLaMA-3B and further extend our findings to Mistral-7B which is evaluated by open-sourced preference model and GPT4. Code available here: https://github.com/avalonstrel/Mitigating-the-Alignment-Tax-of-RLHF.git.

cs.LG↗
arXiv2024

Semi-Supervised Reward Modeling via Iterative Self-Training

Reward models (RM) capture the values and preferences of humans and play a central role in Reinforcement Learning with Human Feedback (RLHF) to align pretrained large language models (LLMs). Traditionally, training these models relies on extensive human-annotated preference data, which poses significant challenges in terms of scalability and cost. To overcome these limitations, we propose Semi-Supervised Reward Modeling (SSRM), an approach that enhances RM training using unlabeled data. Given an unlabeled dataset, SSRM involves three key iterative steps: pseudo-labeling unlabeled examples, selecting high-confidence examples through a confidence threshold, and supervised finetuning on the refined dataset. Across extensive experiments on various model configurations, we demonstrate that SSRM significantly improves reward models without incurring additional labeling costs. Notably, SSRM can achieve performance comparable to models trained entirely on labeled data of equivalent volumes. Overall, SSRM substantially reduces the dependency on large volumes of human-annotated data, thereby decreasing the overall cost and time involved in training effective reward models.

cs.LG↗
arXiv2024

Interpretable Preferences via Multi-Objective Reward Modeling and Mixture-of-Experts

Reinforcement learning from human feedback (RLHF) has emerged as the primary method for aligning large language models (LLMs) with human preferences. The RLHF process typically starts by training a reward model (RM) using human preference data. Conventional RMs are trained on pairwise responses to the same user request, with relative ratings indicating which response humans prefer. The trained RM serves as a proxy for human preferences. However, due to the black-box nature of RMs, their outputs lack interpretability, as humans cannot intuitively understand why an RM thinks a response is good or not. As RMs act as human preference proxies, we believe they should be human-interpretable to ensure that their internal decision processes are consistent with human preferences and to prevent reward hacking in LLM alignment. To build RMs with interpretable preferences, we propose a two-stage approach: i) train an Absolute-Rating Multi-Objective Reward Model (ArmoRM) with multi-dimensional absolute-rating data, each dimension corresponding to a human-interpretable objective (e.g., honesty, verbosity, safety); ii) employ a Mixture-of-Experts (MoE) strategy with a gating network that automatically selects the most suitable reward objectives based on the context. We efficiently trained an ArmoRM with Llama-3 8B and a gating network consisting of a shallow MLP on top of the ArmoRM. Our trained model, ArmoRM-Llama3-8B, obtains state-of-the-art performance on RewardBench, a benchmark evaluating RMs for language modeling. Notably, the performance of our model surpasses the LLM-as-a-judge method with GPT-4 judges by a margin, and approaches the performance of the much larger Nemotron-4 340B reward model.

cs.LG↗
arXiv2024

SAM-CLIP: Merging Vision Foundation Models towards Semantic and Spatial Understanding

The landscape of publicly available vision foundation models (VFMs), such as CLIP and Segment Anything Model (SAM), is expanding rapidly. VFMs are endowed with distinct capabilities stemming from their pre-training objectives. For instance, CLIP excels in semantic understanding, while SAM specializes in spatial understanding for segmentation. In this work, we introduce a simple recipe to efficiently merge VFMs into a unified model that absorbs their expertise. Our method integrates techniques of multi-task learning, continual learning, and distillation. Further, it demands significantly less computational cost compared to traditional multi-task training from scratch, and it only needs a small fraction of the pre-training datasets that were initially used to train individual models. By applying our method to SAM and CLIP, we obtain SAM-CLIP: a unified model that combines the capabilities of SAM and CLIP into a single vision transformer. Compared with deploying SAM and CLIP independently, our merged model, SAM-CLIP, reduces storage and compute costs for inference, making it well-suited for edge device applications. We show that SAM-CLIP not only retains the foundational strengths of SAM and CLIP, but also introduces synergistic functionalities, notably in zero-shot semantic segmentation, where SAM-CLIP establishes new state-of-the-art results on 5 benchmarks. It outperforms previous models that are specifically designed for this task by a large margin, including +6.8% and +5.9% mean IoU improvement on Pascal-VOC and COCO-Stuff datasets, respectively.

cs.CV↗
arXiv2024

Enhancing Compositional Generalization via Compositional Feature Alignment

Real-world applications of machine learning models often confront data distribution shifts, wherein discrepancies exist between the training and test data distributions. In the common multi-domain multi-class setup, as the number of classes and domains scales up, it becomes infeasible to gather training data for every domain-class combination. This challenge naturally leads the quest for models with Compositional Generalization (CG) ability, where models can generalize to unseen domain-class combinations. To delve into the CG challenge, we develop CG-Bench, a suite of CG benchmarks derived from existing real-world image datasets, and observe that the prevalent pretraining-finetuning paradigm on foundational models, such as CLIP and DINOv2, struggles with the challenge. To address this challenge, we propose Compositional Feature Alignment (CFA), a simple two-stage finetuning technique that i) learns two orthogonal linear heads on a pretrained encoder with respect to class and domain labels, and ii) fine-tunes the encoder with the newly learned head frozen. We theoretically and empirically justify that CFA encourages compositional feature learning of pretrained models. We further conduct extensive experiments on CG-Bench for CLIP and DINOv2, two powerful pretrained vision foundation models. Experiment results show that CFA outperforms common finetuning techniques in compositional generalization, corroborating CFA's efficacy in compositional feature learning.

cs.CV↗
arXiv2024

Den-SOFT: Dense Space-Oriented Light Field DataseT for 6-DOF Immersive Experience

We have built a custom mobile multi-camera large-space dense light field capture system, which provides a series of high-quality and sufficiently dense light field images for various scenarios. Our aim is to contribute to the development of popular 3D scene reconstruction algorithms such as IBRnet, NeRF, and 3D Gaussian splitting. More importantly, the collected dataset, which is much denser than existing datasets, may also inspire space-oriented light field reconstruction, which is potentially different from object-centric 3D reconstruction, for immersive VR/AR experiences. We utilized a total of 40 GoPro 10 cameras, capturing images of 5k resolution. The number of photos captured for each scene is no less than 1000, and the average density (view number within a unit sphere) is 134.68. It is also worth noting that our system is capable of efficiently capturing large outdoor scenes. Addressing the current lack of large-space and dense light field datasets, we made efforts to include elements such as sky, reflections, lights and shadows that are of interest to researchers in the field of 3D reconstruction during the data capture process. Finally, we validated the effectiveness of our provided dataset on three popular algorithms and also integrated the reconstructed 3DGS results into the Unity engine, demonstrating the potential of utilizing our datasets to enhance the realism of virtual reality (VR) and create feasible interactive spaces. The dataset is available at our project website.

cs.CV↗
arXiv2024

Arithmetic Control of LLMs for Diverse User Preferences: Directional Preference Alignment with Multi-Objective Rewards

Fine-grained control over large language models (LLMs) remains a significant challenge, hindering their adaptability to diverse user needs. While Reinforcement Learning from Human Feedback (RLHF) shows promise in aligning LLMs, its reliance on scalar rewards often limits its ability to capture diverse user preferences in real-world applications. To address this limitation, we introduce the Directional Preference Alignment (DPA) framework. Unlike the scalar-reward RLHF, DPA incorporates multi-objective reward modeling to represent diverse preference profiles. Additionally, DPA models user preferences as directions (i.e., unit vectors) in the reward space to achieve user-dependent preference control. Our method involves training a multi-objective reward model and then fine-tuning the LLM with a preference-conditioned variant of Rejection Sampling Finetuning (RSF), an RLHF method adopted by Llama 2. This method enjoys a better performance trade-off across various reward objectives. In comparison with the scalar-reward RLHF, DPA offers users intuitive control over LLM generation: they can arithmetically specify their desired trade-offs (e.g., more helpfulness with less verbosity). We also validate the effectiveness of DPA with real-world alignment experiments on Mistral-7B. Our method provides straightforward arithmetic control over the trade-off between helpfulness and verbosity while maintaining competitive performance with strong baselines such as Direct Preference Optimization (DPO).

cs.LG↗
arXiv2024

Predicting Properties of Quantum Systems with Conditional Generative Models

Machine learning has emerged recently as a powerful tool for predicting properties of quantum many-body systems. For many ground states of gapped Hamiltonians, generative models can learn from measurements of a single quantum state to reconstruct the state accurately enough to predict local observables. Alternatively, classification and regression models can predict local observables by learning from measurements on different but related states. In this work, we combine the benefits of both approaches and propose the use of conditional generative models to simultaneously represent a family of states, learning shared structures of different quantum states from measurements. The trained model enables us to predict arbitrary local properties of ground states, even for states not included in the training data, without necessitating further training for new observables. We first numerically validate our approach on 2D random Heisenberg models using simulations of up to 45 qubits. Furthermore, we conduct quantum simulations on a neutral-atom quantum computer and demonstrate that our method can accurately predict the quantum phases of square lattices of 13$\times$13 Rydberg atoms.

quant-ph↗
arXiv2023

Invariant-Feature Subspace Recovery: A New Class of Provable Domain Generalization Algorithms

Domain generalization asks for models trained over a set of training environments to generalize well in unseen test environments. Recently, a series of algorithms such as Invariant Risk Minimization (IRM) have been proposed for domain generalization. However, Rosenfeld et al. (2021) shows that in a simple linear data model, even if non-convexity issues are ignored, IRM and its extensions cannot generalize to unseen environments with less than $d_s+1$ training environments, where $d_s$ is the dimension of the spurious-feature subspace. In this work, we propose Invariant-feature Subspace Recovery (ISR): a new class of algorithms to achieve provable domain generalization across the settings of classification and regression problems. First, in the binary classification setup of Rosenfeld et al. (2021), we show that our first algorithm, ISR-Mean, can identify the subspace spanned by invariant features from the first-order moments of the class-conditional distributions, and achieve provable domain generalization with $d_s+1$ training environments. Our second algorithm, ISR-Cov, further reduces the required number of training environments to $O(1)$ using the information of second-order moments. Notably, unlike IRM, our algorithms bypass non-convexity issues and enjoy global convergence guarantees. Next, we extend ISR-Mean to the more general setting of multi-class classification and propose ISR-Multiclass, which leverages class information and provably recovers the invariant-feature subspace with $\lceil d_s/k\rceil+1$ training environments for $k$-class classification. Finally, for regression problems, we propose ISR-Regression that can identify the invariant-feature subspace with $d_s+1$ training environments. Empirically, we demonstrate the superior performance of our ISRs on synthetic benchmarks. Further, ISR can be used as post-processing methods for feature extractors such as neural nets.

cs.LG↗
arXiv2023

ImmersiveNeRF: Hybrid Radiance Fields for Unbounded Immersive Light Field Reconstruction

This paper proposes a hybrid radiance field representation for unbounded immersive light field reconstruction which supports high-quality rendering and aggressive view extrapolation. The key idea is to first formally separate the foreground and the background and then adaptively balance learning of them during the training process. To fulfill this goal, we represent the foreground and background as two separate radiance fields with two different spatial mapping strategies. We further propose an adaptive sampling strategy and a segmentation regularizer for more clear segmentation and robust convergence. Finally, we contribute a novel immersive light field dataset, named THUImmersive, with the potential to achieve much larger space 6DoF immersive rendering effects compared with existing datasets, by capturing multiple neighboring viewpoints for the same scene, to stimulate the research and AR/VR applications in the immersive light field domain. Extensive experiments demonstrate the strong performance of our method for unbounded immersive light field reconstruction.

cs.CV↗
arXiv2023

Federated Learning with Classifier Shift for Class Imbalance

Federated learning aims to learn a global model collaboratively while the training data belongs to different clients and is not allowed to be exchanged. However, the statistical heterogeneity challenge on non-IID data, such as class imbalance in classification, will cause client drift and significantly reduce the performance of the global model. This paper proposes a simple and effective approach named FedShift which adds the shift on the classifier output during the local training phase to alleviate the negative impact of class imbalance. We theoretically prove that the classifier shift in FedShift can make the local optimum consistent with the global optimum and ensure the convergence of the algorithm. Moreover, our experiments indicate that FedShift significantly outperforms the other state-of-the-art federated learning approaches on various datasets regarding accuracy and communication efficiency.

cs.LG↗
arXiv2022

AuE-IPA: An AU Engagement Based Infant Pain Assessment Method

Recent studies have found that pain in infancy has a significant impact on infant development, including psychological problems, possible brain injury, and pain sensitivity in adulthood. However, due to the lack of specialists and the fact that infants are unable to express verbally their experience of pain, it is difficult to assess infant pain. Most existing infant pain assessment systems directly apply adult methods to infants ignoring the differences between infant expressions and adult expressions. Meanwhile, as the study of facial action coding system continues to advance, the use of action units (AUs) opens up new possibilities for expression recognition and pain assessment. In this paper, a novel AuE-IPA method is proposed for assessing infant pain by leveraging different engagement levels of AUs. First, different engagement levels of AUs in infant pain are revealed, by analyzing the class activation map of an end-to-end pain assessment model. The intensities of top-engaged AUs are then used in a regression model for achieving automatic infant pain assessment. The model proposed is trained and experimented on YouTube Immunization dataset, YouTube Blood Test dataset, and iCOPEVid dataset. The experimental results show that our AuE-IPA method is more applicable to infants and possesses stronger generalization ability than end-to-end assessment model and the classic PSPI metric.

cs.LG↗
arXiv2022

Attention Based Relation Network for Facial Action Units Recognition

Facial action unit (AU) recognition is essential to facial expression analysis. Since there are highly positive or negative correlations between AUs, some existing AU recognition works have focused on modeling AU relations. However, previous relationship-based approaches typically embed predefined rules into their models and ignore the impact of various AU relations in different crowds. In this paper, we propose a novel Attention Based Relation Network (ABRNet) for AU recognition, which can automatically capture AU relations without unnecessary or even disturbing predefined rules. ABRNet uses several relation learning layers to automatically capture different AU relations. The learned AU relation features are then fed into a self-attention fusion module, which aims to refine individual AU features with attention weights to enhance the feature robustness. Furthermore, we propose an AU relation dropout strategy and AU relation loss (AUR-Loss) to better model AU relations, which can further improve AU recognition. Extensive experiments show that our approach achieves state-of-the-art performance on the DISFA and DISFA+ datasets.

cs.CV↗
arXiv2022

Future Gradient Descent for Adapting the Temporal Shifting Data Distribution in Online Recommendation Systems

One of the key challenges of learning an online recommendation model is the temporal domain shift, which causes the mismatch between the training and testing data distribution and hence domain generalization error. To overcome, we propose to learn a meta future gradient generator that forecasts the gradient information of the future data distribution for training so that the recommendation model can be trained as if we were able to look ahead at the future of its deployment. Compared with Batch Update, a widely used paradigm, our theory suggests that the proposed algorithm achieves smaller temporal domain generalization error measured by a gradient variation term in a local regret. We demonstrate the empirical advantage by comparing with various representative baselines.

cs.LG↗
arXiv2022

Asynchronous Training Schemes in Distributed Learning with Time Delay

In the context of distributed deep learning, the issue of stale weights or gradients could result in poor algorithmic performance. This issue is usually tackled by delay tolerant algorithms with some mild assumptions on the objective functions and step sizes. In this paper, we propose a different approach to develop a new algorithm, called $\textbf{P}$redicting $\textbf{C}$lipping $\textbf{A}$synchronous $\textbf{S}$tochastic $\textbf{G}$radient $\textbf{D}$escent (aka, PC-ASGD). Specifically, PC-ASGD has two steps - the $\textit{predicting step}$ leverages the gradient prediction using Taylor expansion to reduce the staleness of the outdated weights while the $\textit{clipping step}$ selectively drops the outdated weights to alleviate their negative effects. A tradeoff parameter is introduced to balance the effects between these two steps. Theoretically, we present the convergence rate considering the effects of delay of the proposed algorithm with constant step size when the smooth objective functions are weakly strongly-convex and nonconvex. One practical variant of PC-ASGD is also proposed by adopting a condition to help with the determination of the tradeoff parameter. For empirical validation, we demonstrate the performance of the algorithm with two deep neural network architectures on two benchmark datasets.

cs.LG↗
arXiv2022

Provable Domain Generalization via Invariant-Feature Subspace Recovery

Domain generalization asks for models trained over a set of training environments to perform well in unseen test environments. Recently, a series of algorithms such as Invariant Risk Minimization (IRM) has been proposed for domain generalization. However, Rosenfeld et al. (2021) shows that in a simple linear data model, even if non-convexity issues are ignored, IRM and its extensions cannot generalize to unseen environments with less than $d_s+1$ training environments, where $d_s$ is the dimension of the spurious-feature subspace. In this paper, we propose to achieve domain generalization with Invariant-feature Subspace Recovery (ISR). Our first algorithm, ISR-Mean, can identify the subspace spanned by invariant features from the first-order moments of the class-conditional distributions, and achieve provable domain generalization with $d_s+1$ training environments under the data model of Rosenfeld et al. (2021). Our second algorithm, ISR-Cov, further reduces the required number of training environments to $O(1)$ using the information of second-order moments. Notably, unlike IRM, our algorithms bypass non-convexity issues and enjoy global convergence guarantees. Empirically, our ISRs can obtain superior performance compared with IRM on synthetic benchmarks. In addition, on three real-world image and text datasets, we show that both ISRs can be used as simple yet effective post-processing methods to improve the worst-case accuracy of (pre-)trained models against spurious correlations and group shifts.

cs.LG↗
arXiv2022

Understanding Gradual Domain Adaptation: Improved Analysis, Optimal Path and Beyond

The vast majority of existing algorithms for unsupervised domain adaptation (UDA) focus on adapting from a labeled source domain to an unlabeled target domain directly in a one-off way. Gradual domain adaptation (GDA), on the other hand, assumes a path of $(T-1)$ unlabeled intermediate domains bridging the source and target, and aims to provide better generalization in the target domain by leveraging the intermediate ones. Under certain assumptions, Kumar et al. (2020) proposed a simple algorithm, Gradual Self-Training, along with a generalization bound in the order of $e^{O(T)} \left(\varepsilon_0+O\left(\sqrt{log(T)/n}\right)\right)$ for the target domain error, where $\varepsilon_0$ is the source domain error and $n$ is the data size of each domain. Due to the exponential factor, this upper bound becomes vacuous when $T$ is only moderately large. In this work, we analyze gradual self-training under more general and relaxed assumptions, and prove a significantly improved generalization bound as $\varepsilon_0+ O \left(TΔ+ T/\sqrt{n}\right) + \widetilde{O}\left(1/\sqrt{nT}\right)$, where $Δ$ is the average distributional distance between consecutive domains. Compared with the existing bound with an exponential dependency on $T$ as a multiplicative factor, our bound only depends on $T$ linearly and additively. Perhaps more interestingly, our result implies the existence of an optimal choice of $T$ that minimizes the generalization error, and it also naturally suggests an optimal way to construct the path of intermediate domains so as to minimize the accumulative path length $TΔ$ between the source and target. To corroborate the implications of our theory, we examine gradual self-training on multiple semi-synthetic and real datasets, which confirms our findings. We believe our insights provide a path forward toward the design of future GDA algorithms.

cs.LG↗