题 目：Learning in Artificial and Biological Systems - From In-Context Learning Phase Transitions to Animal behavioral learning

主 讲：崔文平 博士后研究员

时 间：2026年3月26日（周四）下午14:30-15:30

地 点：琳恩图书馆111报告厅

嘉宾介绍：

崔文平，现为普林斯顿大学博士后研究员。2011年在中国科学技术大学天体物理学专业取得学士学位，2014年在德国波恩大学取得物理学硕士学位，2021年在美国波士顿学院取得物理学博士学位。2021-2024年期间在加州大学圣芭芭拉分校-卡维里理论物理研究所从事博士后研究。研究领域为生物物理，特别是利用统计物理来探索复杂系统的基本原理。研究方向包括生态与演化，细胞感知与响应机制，动物行为学，以及人工神经网络的注意力学习机制。

报告摘要：

How do artificial and biological systems learn structure from their environments? I present two projects that explore shared principles of learning across these domains.

In artificial systems, we study transformers, which exhibit in-context learning: the ability to infer task-specific structure from limited input data without updating model parameters. We provide a mechanistic and dynamical characterization of in-context generalization in a transformer trained on discrete stationary Markov chains. Training gives rise to two distinct algorithmic phases, a unigram phase and a bigram phase, with the bigram solution implemented by a statistical induction head. We derive an effective theory for the learning dynamics of this induction head, explain its abrupt emergence, and show that the transition time is controlled by statistical biases in the data.

In biological systems, we investigate how rats learn continuous pose templates through reinforcement learning. Using an analysis pipeline for high-dimensional trajectory data, we show that behavior can be decomposed into reusable motion motifs, most of which appear to be innate. Our results suggest that biological motor learning follows a select-and-refine strategy, in which innate motifs are recruited and refined through reinforcement rather than generated from scratch.