A comparison of representational learning dynamics in humans and deep neural networks

Semantic knowledge - the ability to acquire, manipulate, and employ conceptual representations - is a crucial component of human cognition. A rich set of empirical investigations have elucidated common patterns in the organisation, acquisition, and deployment of semantic knowledge in humans. Most recently neural network models of semantic cognition have attempted to explain the neural implementation of these rich representations. Crucially for the current proposal, a mathematical analysis of deep linear neural networks trained for semantic cognition has allowed predictions of how representations are formed and modified over the course of learning. The proposed project seeks to understand if the representational evolution observed in these models will also be reflected in the neural activity of humans. While most previous research on the neural substrate of semantic cognition examined representation after acquisition, this project seeks to record neural representations over the course of learning. In doing so, we will find out if several key phenomena in the learning dynamics of neural networks such as progressive hierarchical differentiation, stage-like transitions, and illusory correlations are also prevalent in the human neural substrate and behaviour. To answer these questions, we will employ three methods: a semantic learning experiment with human participants, longitudinal tracking of representational patterns using multivariate analysis of fMRI data, and the training of simple deep neural networks on the same task as our human participants. The project will allow us to examine the validity of neural network models as proxies for human semantic cognition. This will be done in a novel manner by tracking learning dynamics over time, thus notably contributing to contemporary research on semantic cognition.