Title: Efficient multi-objective optimization of neural dynamical systems through joint learning of objectives, constraints and sensitivities
Abstract: Biophysical models of neurons and neuronal networks are described by dynamic equations with numerous parameters that often cannot be directly determined experimentally. Instead, they must be constrained with optimization techniques that iterate through the parameter space in order to determine which combinations of parameters recapitulate experimentally measured features of neural activity. These optimization techniques are time consuming and incur large computational costs. In this talk, we present a novel multi-objective optimization framework that combines several approaches, namely surrogate modeling via non-parametric regression methods, constrained sampling, sensitivity-guided evolutionary optimization, and distributed parallel computing, in order to estimate a surrogate model with a minimal number of computationally expensive model evaluations. We demonstrate that our optimization approach can be used with neuron models based on the Pinsky-Rinzel formalism to quantitatively reproduce key experimentally-obtained electrophysiological features of several major neuron classes as well as network-scale models of the Mind In Vitro platform. We discuss how this optimization framework can be used to develop increasingly sophisticated computing applications that capture key features of in-vitro neural networks.
Bio: Frithjof Gressman is a PhD Candidate in the Siebel School of Computing and Data Science at the University of Illinois, advised by Prof. Lawrence Rauchwerger. This talk is co-presented by Dr. Ivan Raikov. Dr. Raikov is a Research Engineer in the department of Neurosurgery at Stanford University, a member of the research group of Prof. Ivan Soltesz, Professor of Neurosurgery and Neurosciences.
Part of the Illinois Computer Science Speakers Series.
Food will be provided after the seminar.
If accommodation is required, please email <communications@cs.illinois.edu>. Someone from our staff will contact you to discuss your specific needs.