数苑博雅讲座（75）：From Biological Neuronal Computation to Artificial Neural Networks
来源：
|
发布时间：2026-08-28
报告时间：
2026年8月30日（周日）下午17:00-18:00
报告地点：
国交二号楼315会议室
报告人：
李松挺
报告人单位：
上海交通大学
邀请人：
张晓飞
摘要：
This talk explores the connection between biological neuronal computation and the design of its inspired novel artificial neural networks. We begin by examining how single neurons integrate spatiotemporal inputs through their dendritic structures. Using mathematical modeling, theoretical analysis, and numerical simulations, we uncover a bilinear integration rule that captures essential features of dendritic information processing. Building on this principle, we introduce a compact and interpretable architecture termed the Dendritic Bilinear Neural Network (DBNN). This model accurately predicts subthreshold membrane potentials and spike timing across diverse neuron types, while exhibiting strong generalization and robustness. Beyond prediction, DBNN reproduces key dendritic computations, including direction selectivity, coincidence detection, and logical operations, with parameters that closely align with biologically observed mechanisms. As a novel neural network architecture, DBNN achieves strong performance on temporal learning tasks, such as dynamic vision and neuromorphic datasets. Overall, this work illustrates how insights from computational neuroscience can guide the development of more efficient, interpretable, and biologically grounded AI models.
来源：
发布时间：2026-08-28
报告时间：
