learning of dynamics

Real-world, non-autonomous dynamical systems are hard to predict. This is a central challenge for artificial and biological neural networks. Despite the underlying non-autonomy, the presence of structure in latent factors can lead to the recurrence of dynamical regimes. A major research goal of DIG is understanding how artificial and biological intelligence can learn to identify such repeated dynamics and leverage such knowledge to make better informed predictions and action plans.

To this end, we are:

• Developing techniques for comparing non-autonomous dynamical systems and using this insight to make predictions (Redman et al., 2025). We test our approaches on real-world data, including forecasting infectious diseases in the US and Canada (Redman & Mullany, 2025).

• Building recurrent neural network (RNN) models for examining predictive behavior and neural computations that emerge during continuous pursuit of a moving target. We closely collaborate with systems neuroscientists (Alexander Lab) who experimentally probe predictions of the models.