Export
Weiming Ou
· 2 previous affiliations
- PhD student · 2023–2027
- MS student · 2021–2023
Research Overview
Weiming Ou focuses on the theoretical analysis of adaptive optimization algorithms and their application to complex training environments, such as min-max games and reinforcement learning. This work centers on establishing convergence guarantees and understanding the dynamical behavior of stochastic optimizers like Adam and AdaGrad. By employing continuous-time modeling and rigorous convergence analysis, Ou investigates how algorithmic parameters influence stability, robustness, and the ability to find generalized solutions in non-convex and multi-player settings.
Research Trajectory
The available research record, concentrated in 2026, indicates a primary focus on the mathematical foundations of adaptive gradient methods. A significant portion of this work is dedicated to bridging the gap between the empirical success of these algorithms and their theoretical understanding, particularly in scenarios that deviate from standard empirical risk minimization.
One major trajectory involves the study of adaptive optimizers in the context of zero-sum and min-max games, such as those encountered in Generative Adversarial Networks (GANs). Ou has contributed to research that challenges the assumption that the properties of Adam in minimization tasks naturally extend to game-theoretic settings. This work utilizes Ordinary Differential Equation (ODE) approaches to derive high-fidelity continuous-time models, revealing that the stability of these algorithms in games requires fundamentally different parameter configurations—specifically regarding first- and second-order momentum—than those used in standard classification or regression.
Parallel to the study of game dynamics, Ou has investigated the fundamental convergence properties of AdaGrad. This line of inquiry focuses on "almost sure" last-iterate convergence rates, a stringent form of stability guarantee that ensures an algorithm converges for almost every individual training trajectory rather than just in expectation. This research addresses gaps in the existing literature regarding the robustness of AdaGrad under arbitrary stopping times in both convex and non-convex settings.
Additionally, Ou has participated in research extending these optimization principles to reinforcement learning (RL). This work addresses the challenges of transitioning from offline pre-training to online fine-tuning, specifically focusing on the trade-offs between optimism and pessimism in model-based RL. This suggests an interest in how optimization dynamics and objective formulations can mitigate the distribution shifts that occur when an agent moves from static datasets to active environment interaction.
Current Directions
The current research focus appears to be divided between the refinement of optimization theory and the practical application of these theories to stabilize complex machine learning models.
Theoretical Analysis of Adaptive Optimizers
A central theme in Ou's recent work is the use of continuous-time analysis to decode the behavior of Adam. By treating the discrete steps of the optimizer as a continuous dynamical system, this research identifies how momentum parameters influence local convergence and implicit gradient regularization. This direction is particularly relevant for understanding why certain hyperparameter settings lead to "flat" solutions, which are often associated with better generalization in deep learning.
Robust Convergence and Stability
The pursuit of robust convergence rates for algorithms like AdaGrad represents a commitment to providing stronger theoretical foundations for the tools used in deep model training. By establishing rates that hold under any stopping time and for the last iterate, this work aims to provide practitioners with more reliable guarantees of stability, reducing the reliance on heuristic-based stopping criteria.
Algorithmic Trade-offs in Reinforcement Learning
In the domain of reinforcement learning, the focus lies on the "optimism-pessimism trade-off." This involves identifying how to balance the need for exploration (optimism) with the need for safety and data-driven constraints (pessimism) during the offline-to-online transition. This research seeks to replace expensive bi-level optimization procedures with more efficient single-objective formulations.
Representative Work
AdaGrad Converges in a Robust Sense: Almost Sure Last-Iterate Rates under Any Stopping Time establishes a rigorous framework for the convergence of AdaGrad. As the first author, Ou led the establishment of almost sure convergence rates for the last-iterate in strongly convex settings and for the best-iterate in non-convex settings. The work is significant for its focus on robustness, providing guarantees that are valid under arbitrary stopping times and flexible gradient dependencies, thereby offering a more complete theoretical picture of one of the most widely used stochastic optimization algorithms.
In Understanding Dynamics of Adam in Zero-Sum Games: An ODE Approach, Ou contributed as the second author to a study that explains the divergent behavior of Adam when applied to games like GANs. The researchers derived a high-fidelity ODE to model the algorithm's continuous-time trajectory, uncovering a "momentum reversal" phenomenon. The study demonstrates that to achieve stability and locate flat solutions in zero-sum games, practitioners must adopt parameter settings—specifically smaller first-order and larger second-order momentum—that are counterintuitive when compared to standard minimization tasks.
Continuous-time Analysis of Adam in Min-Max Games: Local Convergence and Implicit Gradient Regularization further explores the dynamics of Adam in multi-player settings. As the second author, Ou participated in investigating how first- and second-order momentum parameters influence local convergence and implicit regularization. The paper refutes the notion that Adam’s minimization properties transfer seamlessly to min-max games, providing a detailed analysis of how the algorithm's internal dynamics change the optimization landscape in these complex scenarios.
Ou also participated in the development of Towards Optimism-Pessimism Trade-off in Model-based Offline-to-Online Reinforcement Learning. This work addresses the sample efficiency of model-based RL by examining the distribution shifts that occur during the transition from offline data to online interaction. The research identifies the optimism-pessimism trade-off as a primary challenge in offline training and proposes a framework to manage this trade-off more efficiently than traditional bi-level optimization methods.
Profile updated
Loading new papers…
Other researchers whose papers cluster into the same research directions — a proxy for overlapping technical interests and shared intellectual neighborhoods.
See something wrong? Report a mistake