BEGIN:VCALENDAR
BEGIN:VEVENT
SUMMARY:EE Seminar
DESCRIPTION:Speaker: Terence D. Sanger, MD PhD, Depts of Biomedical Engineering, Biokinesiology, and Child Neurology/USC
Talk Title: A Bayesian nonlinear filter and a stochastic nonlinear control algorithm suitable for estimation and control by populations of spiking neurons
Abstract: The best-known examples of Bayesian nonlinear filters are the Kushner and Zakai equations which unfortunately have limited applicability to important classes of real-world problems. I derive a general nonlinear filter with broad applicability that can be shown to integrate to Bayes' rule over short time intervals. The filter extracts maximal information per unit time, in the sense that the rate of decrease of the entropy of the estimate is equal to the mutual information between the state and the observation. I show that this filter has a straightforward parallel implementation, and I show an efficient representation using Poisson-distributed spiking neurons.\n
\n
I then show that this technique can be extended to a class of stochastic nonlinear controllers. These controllers extend linear feedback controllers and permit control of systems with non-Gaussian noise or state uncertainty, asymmetric cost or perturbations, or state measurements that are not characterized by additive Gaussian noise. The theory is based on Stochastic Dynamic Operators (SDOs) in which the fundamental signals used for feedback are not estimates of state, but estimates of the probability distribution of state. This allows control to vary depending on the degree of state uncertainty (eg: one might drive more slowly if visibility is poor). The reference signal used for control is not a desired time-varying reference state, but a time-varying cost function that assign a value to every potential state. Such cost functions can represent asymmetric penalties and discontinuities in cost (eg: a cliff to one side of a road). Feedback control uses Bayesian statistics to combine the uncertain state estimate (from a nonlinear filter) and the time-varying cost function to produce an estimated motor command. The command is the solution to a short-term optimization problem. As with the Bayesian nonlinear filter, populations of spiking neurons provide a good representation for SDOs and an efficient control algorithm. I will show a real-time implementation of a feedback controller for a desktop robot arm using a population of 900 simulated spiking neurons that tracks the desired minimum cost and stably resists perturbations.
Biography: Terry Sanger holds an SM in Applied mathematics (Harvard), PhD in Electrical Engineering and Computer Science (MIT), and MD (Harvard), with medical specialization in Child Neurology and Movement Disorders. He is currently Associate Professor of Biomedical Engineering, Neurology, and Biokinesiology, and he is the director of the Pediatric Movement Disorders Clinic at Childrens Hospital of Los Angeles, and the Health Technology and Engineering program at USC (HTE@USC).\n
\n
His research on disorders of developmental motor control is driven by his interest in finding new treatments for children with movement disorders including dystonia, chorea, spasticity, and dyspraxia. He has a particular interest in computational motor learning, and the role of motor learning in recovery from childhood brain injury. Major focus areas of laboratory research include wearable devices to promote motor learning, EMG-driven communication devices and assistive prosthetics, and modeling of the electrophysiology of deep-brain stimulation. Personal involvement in motor control and motor learning includes snowboarding, jazz and classical piano, bluegrass banjo, and ballroom dance with particular focus on Argentine Tango.
Host: Professor Sandeep K. Gupta, sandeep@usc.edu
DTSTART:20160915T153000
LOCATION:EEB 248
URL;VALUE=URI:
DTEND:20160915T163000
END:VEVENT
END:VCALENDAR