dynamics of learning

A defining feature of learning in artificial and biological neural networks is that it is a dynamic process. However, it has historically been challenging to utilize the rich set of tools developed in dynamical systems theory to characterize learning, as the equations governing learning are complex and (in some cases) not known. A central research goal of DIG is developing and leveraging data-driven dynamical systems theory methods to shed light on the complex emergent phenomena associated with artificial and biological learning.

To this end, we are:

• Developing techniques to compare the learning dynamics associated with the training of different deep neural networks (DNNs) (Redman et al., 2024). We closely collaborate with academic (UCSB, Johns Hopkins University) and industry (AIMdyn Inc.) researchers who are at the forefront of advancing Koopman operator theory.