Brain-inspired dual-pathway neural network architecture and its generalization analysis

Abstract

In this study, we explored the neural mechanism of global topological perception in the human visual system. We showed strong evidence that the retinotectal pathway in the archicortex of the human brain is responsible for global topological perception, and for modulating the local feature processing in the classical ventral visual pathway. Inspired by this recent cognitive discovery, we developed a novel CogNet architecture to emulate the global-local dichotomy of human visual cognitive mechanisms. The thorough experimental results indicate that the proposed CogNet not only significantly improves image classification accuracies but also effectively addresses the texture bias problem observed in baseline CNN models. We have also conducted mathematical analysis for the generalization gap for general neural networks. Our theoretical derivations suggest that the Hurst parameter, a measure of the curvature of the loss landscape, can closely bind the generalization gap. A larger Hurst parameter corresponds to a better generalization ability. We found that our proposed CogNet achieves a lower test error and attains a larger Hurst parameter, strengthening its superiority over the baseline CNN models further.