Training recurrent neural networks (RNNs) has revolutionized the way how systems neuroscientists form hypothesis when studying circuit mechanisms in various problems. However, the trained RNNs oftentimes are difficult to be interpreted, inconsistent with neural data or not necessarily comprising the full set of biological solutions. In this talk, I will introduce an interpretable RNN training framework, namely Restricted-RNN, capable of generating interpretable circuit hypothesis through a proposing-and-testing procedure that seamlessly integrates multilevel descriptions ranging from computational-, collective- to implementational-level ones. The validity of Restricted-RNN was demonstrated through the identification of novel data-compatible circuit mechanisms in a variety of macaque cognitive tasks, including parametric working memory, sequence working memory and perceptual decision-making. The key predictions derived from this modeling approach were confirmed by monkey prefrontal and parietal neurophysiological data. Critically, the interpretable nature of Restricted-RNN endowed us a unified theory to explain the seemingly disparate phenomena across different tasks with a novel neural control state space, providing an intriguing geometric understanding for the ubiquitous control in cognitive processes.

Biography

闵斌，临港实验室青年研究员，博士生导师。2004-2013年就读于北京大学数学科学学院，先后获得理学学士学位和理学博士学位，2013-2018年于纽约大学从事博士后研究工作，2018-2022年于上海脑科学与类脑研究中心担任副研究员，2022年9月起于临港实验室任青年研究员。主要研究方向为计算认知神经科学，以神经数据分析与计算建模（特别是新型可解释神经网络建模）为研究手段，致力于探寻大脑神经网络的一般计算原理及其在脑机接口领域的应用。相关研究工作以通讯/共同通讯作者身份发表在Science (2022, 2024), Neuron (2024)，eLife (2024)等期刊上，并获评2022年度“中国神经科学重大进展”。

Inviter: Hai-Jun Zhou