Representations in Biological and Artificial Neural Networks

Abstract

Remarkably, artificial neural networks (ANNs) have shown astounding success in almost all aspects of artificial intelligence. Meanwhile, large scale experiments have gathered an unprecedented amount of data about the biological brain, both anatomical and functional. In this thesis, we make a series of interconnected endeavors to link ANNs with biological brains, and show that such links can shed light on our understanding of both systems. First, we establish and validate a paradigm for comparing ANN models with large scale functional data sets from mouse visual cortex. We show that comparing ANNs to the real brain is not only possible to do in a reliable way, but also helpful in revealing insights about computation in the biological brain. Second, we present the first (to our knowledge) ANN model of the mouse visual cortex (MouseNet) that is constrained by large-scale mesoscopic anatomical data. With the MouseNet model, we demonstrate the computational capabilities of a mouse- sized architecture and quantify the extent to which it recapitulates the neural representation of images in mouse visual cortex. Finally, we utilize the mathematical framework of linear ANNs to study learning dynamics in a simple model with parallel pathways, an important network feature that appears in MouseNet and is common in biological brains. We examine and quantify the surprisingly rich dynamics by which the learning process distributes the task-related knowledge among different network pathways.