Principles of deep networks and their relevance for the neural circuit mechanisms underlying learning

How the brain learns remains much of a mystery. We are often surprised by our rapid learning on some tasks and inability on others, the diversity across individuals and the seemingly unpredictable course of learning trajectories. Remarkably, the advent of deep learning has revealed that artificial neural networks (ANNs) can flexibly learn complex tasks and share many similarities with the brain. As a growing community of theorists uncover the principles governing learning in ANNs, this presents a unique opportunity to investigate whether similar principles underlie learning in the brain. Here, I propose to experimentally test deep learning as a framework with which to model, measure and make predictions about neural learning mechanisms. To do so, I have developed behavioural tasks for rodents that probe key learning phenomena: long-term learning through stages, the transfer or forgetting of previous knowledge in continual learning and the learning of non-linear computations. I aim to apply large-scale recording techniques to measure the neuronal teaching signals and changes in activity throughout learning, and compare these to predictions from deep learning models. This work has the potential to transform our understanding of learning in the brain by constructing a formal framework to describe learning across neural circuits.