Event
Communication, Control and Signal Processing Seminar: Rajesh Sundaresan, "Detecting oddball targets"
Wednesday, August 29, 2018
5:00 p.m.-6:30 p.m.
2168 AV Williams Building
Ajaykrishnan Nageswaran
301 405 3661
ajayk@umd.edu
http://www.ece.umd.edu/seminars/ccsp/
Communication, Control and Signal Processing Seminar
Learning to Detect an Oddball Target
Rajesh Sundaresan
Indian Institute of Science
Abstract
What policy should you employ if you were to locate an oddball target among many (two or more) distracters? You do not know the statistics of the observations that a 'view' of the oddball produces. Nor do you know the statistics of the observations arising from a view of a distractor. The objective is to minimize the search time, subject to keeping the probability of false detection low. Clearly, the common statistics of the distracter-generated observations should be compared, recognized as similar, and further recognized as different from those generated by the oddball target. But where should you look next, when should you stop, and when you stop, what should you decide as the oddball location?
When the observations are points of a Poisson process (as in firings of neurons), and when the probability of false detection is driven down to zero, we will discuss the best growth rate of the expected search time. We will also identify an asymptotically optimal search policy that achieves the best growth rate. Finally, we will apply our results to a particular visual search experiment studied recently by neuroscientists and will highlight how prior information affects search performance.