CE⁴L: Continual Ego, Exo, and Ego-Exo Learning
Introduces a multi-view continual learning benchmark spanning four video tasks and VISTA, a lightweight adapter method with training-free subspace routing for simultaneous task and viewpoint shifts.
Hongwei Yan 燕鸿伟
PhD student · School of Life Sciences, Tsinghua University
I am jointly advised by Yi Zhong and Kaisheng Ma. I also collaborate with Liyuan Wang in Tsinghua's Department of Psychological and Cognitive Sciences and am a research intern at the Beijing Academy of Artificial Intelligence (BAAI).
I trained in computer science in Tsinghua's Yao Class before starting my PhD in biology. This background shapes both sides of my research: translating mechanisms of learning and memory into machine-learning algorithms, and developing computational methods to study neural and biological data.
Research
- Brain-inspired AI. I explore how foundation models and agents can accumulate knowledge, reason, adapt over time, and improve through experience. My research draws on biological learning and memory, with published work in continual learning and generalization.
- AI for Science. I build foundation models from neural and biological data and develop AI tools for neuroscience and computational biology, spanning multimodal representations, transcriptomics, and scientific data acquisition and analysis.
Selected publications
All publications* Equal contribution.
FlyPrompt: Brain-Inspired Random-Expanded Routing with Temporal-Ensemble Experts for General Continual Learning
Combines brain-inspired random-expanded routing with temporal-ensemble experts to learn from single-pass, evolving data streams without explicit task boundaries.
Right Time to Learn: Promoting Generalization via Bio-inspired Spacing Effect in Knowledge Distillation
Introduces a time interval between teacher and student updates in online and self-distillation, connecting improved generalization to flatter optimization landscapes.
Orchestrate Latent Expertise: Advancing Online Continual Learning with Multi-Level Supervision and Reverse Self-Distillation
Coordinates experts at multiple network depths with multi-level supervision and reverse self-distillation to balance underfitting and overfitting in online continual learning.