Dynamics-inspired Neuromorphic Visual Representation Learning

Zhengqi Pei, Shuhui Wang

Abstract: This paper investigates the dynamics-inspired neuromorphic architecture for neural representation and learning following Hamilton's principle. The proposed approach converts weight-based neural structure to its dynamics-based form that consists of finite sub-models, whose mutual relations measured by computing path integrals amongst their dynamic states are equivalent to the typical neural weights. Based on the entropy reduction process derived from the Euler-Lagrange equations, the feedback signals interpreted as stress forces amongst sub-models push them to move. We first train a dynamics-based neural model from scratch and observe that this model outperforms its corresponding neural models on MNIST. We then convert several pre-trained neural structures into dynamics-based forms, followed by fine-tuning via entropy reduction to obtain the stabilized dynamic states. We observe consistent improvements in these transformed models on ImageNet and WebVision regarding computational complexity, parameter size, testing accuracy, and robustness. Besides, we show the correlation between model performance and structural entropy, providing a new insight into neuromorphic learning.