Yimin Hou

Beihang University
1 previous affiliation
3Machine learning papers tracked by Lacuna
Is this you?

Research Overview

Yimin Hou focuses on deep learning architectures for multilabel classification, low-shot learning, and data-driven modeling of physical systems. Early work concentrated on improving neural network inference and uncertainty assessment for human-in-the-loop systems. More recently, this trajectory has expanded to include collaborative efforts in applying reinforcement learning and bi-level optimization to discover governing equations for complex nonlinear dynamics.

Research Trajectory

In 2021, Hou’s research primarily addressed structural and algorithmic improvements in deep learning, with a specific emphasis on multilabel classification and low-shot learning scenarios. As the first author of Gated Value Network for Multilabel Classification, Hou developed a framework that builds upon deep value networks by directly optimizing the compatibility metric between data and potential labels. This work introduced a feedforward predictor to simplify the complex variable optimization steps typically required during inference. Additionally, it incorporated a gating mechanism designed to filter out confounding factors from input data, enabling more precise end-to-end training for multilabel prediction.

During the same period, Hou also contributed to methodologies for handling limited observational data. As the second author of Human-in-the-Loop Low-Shot Learning, Hou helped design a system that integrates an uncertainty assessment module into low-shot learning pipelines. This approach specifically targets the emergence of out-of-distribution samples during the testing phase. By detecting highly heterogeneous data points that diverge from existing labeled data, the system flags these uncertain samples for active labeling by human annotators, thereby maintaining model reliability in unpredictable environments.

By 2025, Hou’s collaborative efforts expanded into the intersection of machine learning and the physical sciences, focusing on the data-driven discovery of system dynamics. Participating as a middle author in Bi-level identification of governing equations for nonlinear physical systems, Hou contributed to the development of the BILLIE framework. This research leverages policy gradient algorithms from reinforcement learning to perform bi-level optimization, simultaneously discovering and validating governing equations. The work aims to mitigate the persistent risk of overfitting when modeling complex, canonical nonlinear environments, demonstrating applications across turbulent flows, three-body systems, and biological velocity models.

Current Directions

The most recent work in the evidence indicates a shift toward applying advanced optimization and reinforcement learning techniques to scientific discovery. Recent collaborative efforts investigate the challenge of extracting governing equations directly from observational data in nonlinear physical systems. This direction emphasizes hierarchical optimization strategies to balance the automated discovery of mathematical models with rigorous validation. By integrating reinforcement learning into the optimization pipeline, these recent approaches aim to prevent overfitting when modeling complex phenomena ranging from fluid dynamics to molecular biology.

Profile updated

See something wrong? Report a mistake