Talk by Prof. Haiping Huang (CSN Virtual Seminar)
We hereby announce the next talk in the 'Computational and Systems Neuroscience Virtual Seminar' in short: 'CSN Virtual Seminar'
Ensemble perspective for understanding temporal credit assignment: adapting network statistics, stochastic plasticity and disentangled low-dimensional neural dynamics
Recurrent neural networks are widely used in processing complex temporal sequences, like nature language processing and brain dynamics. It remains elusive which neural connections and how weight uncertainty impact the network behavior. In this talk, I will introduce an ensemble method to understand this temporal credit assignment. Our method could precisely identify the critical neural connections, producing an ensemble of candidate networks, which traditional methods could not achieve. Moreover, by theoretical arguments, our statistical model links the network statistics, the emergent neural selectivity, symmetry breaking, stochastic plasticity and low-dimensional learning manifold. Our method can thus be used as a promising tool to explore internal dynamics of widely-used recurrent neural networks.
Reference: W. Zou, C. Li, H. Huang, arXiv:2102.03740
Prof. Haiping Huang
Sun Yat-sen University, China