Dynamics of representational learning in brain and artificial neural networks*

Abstract

Learning happens in realistic neural networks in brains as well as in artificial neural networks (ANN) such as those in deep learning, which has achieved near or above human level performance for an increasing list of specific tasks. However, both their network architectures and the underlying learning rules are significantly different. Realistic neural networks in brains have recurrent connections between neurons while ANN in deep learning has a simple feedforward architecture. Brains learn through updates of the synaptic weights through a local learning rule such as the Hebbian learning rule while the weight parameters in deep learning ANN models are updated to minimize a global loss function. In this talk, we will discuss the commonalities and differences between learning dynamics of realistic neural networks and artificial neural networks in the context of representational learning by using two examples from the mammalian olfactory system: 1) representational drift in piriform cortex; 2) alignment of neural representations from two sides of the brain. *NIH grant R35GM131734