From Explainable AI to Neuroscience: Revealing the Emergence of Computations from the Collective Dynamics of Interacting Neurons

Event Details

Type

Special Seminar Series

Speaker(s)

Surya Ganguli, Ph.D., associate professor, Stanford University

Speaker bio(s)

Concomitant advances in experimental neuroscience and deep learning now enable us to simultaneously observe the activity of many neurons, and then quantitatively describe their dynamics with highly accurate but complex models. But are we then merely replacing something we don’t understand (the brain) with something else we don’t understand (our complex model of it)? How can we instead derive a conceptual understanding of how important computations emerge from the model, as well as derive incisive experimental tests of the model? We will show how to use ideas from explainable AI and applied mathematics to address these questions in a range of systems from vision to navigation to AI systems. In particular, we will explain how the first steps of vision unfold in a single retinal circuit model that can capture over 2 decades of seminal experiments, including natural scene responses. Also we will explain simple principles governing how neural circuits can fuse information from landmarks and self-motion to create stable spatial maps of the world after a single exposure to a novel environment; our explainable model can predict detailed properties of entorhinal grid cell representations in new environments before a mouse even enters them.

Open to

Tri-Institutional