马克斯·普朗克生物控制论研究所Peter Dayan课题组的最新研究探明了无限隐马尔可夫模型可以剖析学习的复杂性。2025年12月30日出版的《自然—神经科学》杂志发表了这项成果。

本文提出了一种动态无限隐半马尔可夫模型，其潜在状态与行为的特定组成部分相关联。该模型可以通过引入新状态来描述新行为，并通过现有状态中的动态捕获更适度的适应。研究组通过将模型拟合到100只小鼠学习对比检测任务的行为数据中来测试该模型。尽管动物在学习这项任务时表现出很大的个体间差异，但大多数小鼠都经历了任务理解的三个阶段，新的行为通常在会话开始时出现，早期反应偏差不能预测后来的反应。课题组研究人员提供了一种新的工具来全面捕捉学习过程中的行为。

据了解，了解一项任务的偶然性是困难的。个人以一种独特的方式学习，在探索和适应的过程中多次修改他们的方法。这些学习曲线的定量表征需要一个既能捕捉新行为又能捕捉现有行为缓慢变化的模型。

附：英文原文

Title: Infinite hidden Markov models can dissect the complexities of learning

Author: Bruijns, Sebastian A., Bougrova, Kcnia, Laranjeira, Ins C., Lau, Petrina Y. P., Meijer, Guido T., Miska, Nathaniel J., Noel, Jean-Paul, Pan-Vazquez, Alejandro, Roth, Noam, Socha, Karolina Z., Urai, Anne E., Dayan, Peter

Issue&Volume: 2025-12-30

Abstract: Learning the contingencies of a task is difficult. Individuals learn in an idiosyncratic manner, revising their approach multiple times as they explore and adapt. Quantitative characterization of these learning curves requires a model that can capture both new behaviors and slow changes in existing ones. Here we suggest a dynamic infinite hidden semi-Markov model, whose latent states are associated with specific components of behavior. This model can describe new behaviors by introducing new states and capture more modest adaptations through dynamics in existing states. We tested the model by fitting it to behavioral data of >100 mice learning a contrast-detection task. Although animals showed large interindividual differences while learning this task, most mice progressed through three stages of task understanding, new behavior often arose at session onset, and early response biases did not predict later ones. We thus provide a new tool for comprehensively capturing behavior during learning.

DOI: 10.1038/s41593-025-02130-x

Source: https://www.nature.com/articles/s41593-025-02130-x