DIEP seminar by Jason Kim
Artificial and biological neural systems produce emergent and generalizable intelligence. Despite significant architectural variability between models and species, the algorithms and computations remain largely the same: when we are hungry, we search our mental model of the world to find food. How do the complex and recurrent dynamics between neural units produce the algorithms and representations that underlie intelligent function? How do we discover them in biological neural recordings?
I will discuss novel methods I am pioneering that combine dynamical systems, differential geometry, and deep learning to tackle these questions. In the first half, I will discuss how to program the interactions of a simple model of neural dynamics, a recurrent neural network, to compile complex algorithms such as addressable memory and computation, virtualization, Boolean logic, and mental simulations of video games. I will discuss how to decompile these algorithms directly from the weights of trained networks. In the second half, I will discuss how to use the differential geometry of autoencoders to quantitatively model and extract the algorithms that mice use to navigate mazes directly from calcium imaging data of thousands of hippocampal CA1 neurons. I will demonstrate that the representational structure of cognitive maps is organized as a small-world graph to simultaneously be locally accurate for reliable action and globally searchable for efficient planning. Combining principles underlying artificial and biological intelligence may provide insights towards building and using generalizable world models.
If you wish to attend this seminar online, please send an email to r.lier@uva.nl to receive the zoom-link.