The ability of artificial neural networks to avoid overfitting is reflected in a certain characteristic structure called “sloppiness”. The signature of sloppiness is that the network has a Fisher Information Metric (FIM) with eigenvalues distributed uniformly on a logarithmic scale. This distribution indicates a large degree of redundancy in the learned parameters; i.e., there is one set of parameters that is tightly constrained by the data, another that can vary twice as much without affecting predictions, and so on. This project aims to expand and formalize our understanding of how sloppiness affects generalization in different artificial neural networks.

In addition, we aim to apply these ideas to understand how human and non-human primates learn tasks and adapt to changes in tasks that require inferences using multi-modal sensory inputs and in complex, dynamic environments, and their benefits are affected by upstream and downstream computation.