Towards Biologically Plausible Deep Neural Networks

Abstract

Human intelligence has long been a source of inspirations for developing artificial intelligence. Computational principles and components discovered in the human brain have been successfully applied to artificial systems. These artificial systems are not only useful for engineering tasks but also serve as computational models of the brain, connecting theories to empirical data. Conversely, artificial intelligence, deep neural networks in particular, has contributed to advancing the understanding of the brain. Deep neural networks when trained adequately can reproduce behavioral and neural data better than previously developed models. Here we present studies that contribute to this interplay between natural and artificial intelligence. We first investigate invariance, a key computational principle that enables robust visual recognition and efficient generalization to new visual concepts, in human vision. Based on the experimental results, we propose deep neural network architectures that support the observed human behavioral properties in invariant recognition tasks. Next, we introduce a comparison framework for deep neural networks, where ground-truth targets are known, such that interpretations from the comparison can be validated. We explore whether deep neural networks with high functional similarity measures can provide reliable insights into the architectural building blocks of the brain.