Session: Robust Machine Learning
Chair: Yuxuan Han
Cluster: Optimization For Data Science
Talk 1: Robust Reinforcement Learning from Corrupted Human Feedback
Speaker: Zixuan Zhang
Abstract: Reinforcement learning from human feedback (RLHF) provides a principled framework for aligning AI systems with human preference data. For various reasons, e.g., personal bias, context ambiguity, lack of training, etc, human annotators may give incorrect or inconsistent preference labels. To tackle this challenge, we propose a robust RLHF approach $R^3M$, which models the potentially corrupted preference label as sparse outliers. Accordingly, we formulate the robust reward learning as an l1-regularized maximum likelihood estimation problem. Computationally, we develop an efficient alternating optimization algorithm, which only incurs negligible computational overhead compared with the standard RLHF approach. Theoretically, we prove that under proper regularity conditions, $R^3M$ can consistently learn the underlying reward and identify outliers, provided that the number of outlier labels scales sublinearly with the preference sample size. Furthermore, we remark that $R^3M$ is versatile and can be extended to various preference optimization methods, including direct preference optimization (DPO). Our experiments on robotic control and natural language generation with large language models (LLMs) show that $R^3M$ improves robustness of the reward against several types of perturbations to the preference data.
Talk 2: On the Convergence of Projected Bures-Wasserstein Gradient Descent under Euclidean Convexity
Speaker: Yuxuan Han
Abstract: The Bures-Wasserstein (BW) gradient descent method has gained considerable attention in various domains, including Gaussian barycenter, matrix recovery and variational inference problems, due to its alignment with the Wasserstein geometry of normal distributions. Despite its popularity, existing convergence analysis are often contingent upon specific loss functions, and the exploration of constrained settings within this framework remains limited. In this work, we make an attempt to bridge this gap by providing a general convergence rate guarantee for BW gradient descent when the Euclidean convexity of the loss and the constraints is assumed. In an effort to advance practical implementations, we also derive a closed-form solution for the projection onto BW distance-constrained sets, which enables the fast implementation of projected BW gradient descent for problems that arise in the constrained barycenter and distributionally robust optimization literature. Experimental results demonstrate significant improvements in computational efficiency and convergence speed, underscoring the efficacy of our method in practical scenarios.
Talk 3: Approximations to worst-case data dropping: unmasking failure modes
Speaker: Jenny Huang
Abstract: A data analyst might worry about generalization if dropping a very small fraction of data points from a study could change its substantive conclusions. Finding the worst-case data subset to drop poses a combinatorial optimization problem. To overcome this intractability, recent works propose using additive approximations, which treat the contribution of a collection of data points as the sum of their individual contributions, and greedy approximations, which iteratively select the point with the highest impact to drop and re-run the data analysis without that point [Broderick et al., 2020, Kuschnig et al., 2021]. We identify that, even in a setting as simple as OLS linear regression, many of these approximations can break down in realistic data arrangements. Several of our examples reflect masking, where one outlier may hide or conceal the effect of another outlier. Based on the failures we identify, we provide recommendations for users and suggest directions for future improvements.
