Seminar by Sheeraz Ahmad
Active Sensing: An Optimal Inference Model
Sheeraz Ahmad
Univ. of California, San Diego
Date: Wednesday, September 12th, 2012
Time: 5PM
Venue: CS101.Abstract:
Sensory inference under conditions of uncertainty is a major problem in both machine learning and computational neuroscience. An important but poorly understood aspect of sensory processing is the role of active sensing: the use of self-motion to selectively process the most rewarding or informative aspects of the environment. Here, we present a Bayes-optimal inference and decision-making framework for active sensing, which directly minimizes a cost function that takes into account behavioral costs such as response delay, error, and effort. Unlike previously proposed algorithms that optimize heuristic objectives such as expected entropy reduction [Butko et al, 2010] or one-step look-ahead accuracy [Najemnik et al, 2005], this optimal policy can account for search duration as well as location, and is sensitive to contextual factors such as the relative importance of time, error, and effort. We implement the optimal policy, along with the two heuristic policies, for an example visual search task, and illustrate how the heuristic policies deviate from optimal performance in various contexts. We show that the discrepancy is especially large when the cost of time and the cost of switching between sensing locations are high. We demonstrate a potential route for overcoming the computational complexity of the optimal algorithm, especially problematic in large real-world applications, by exploiting the concavity and smoothness of the value function. We show that a basis function approximation to the value function, several orders of magnitude reduced in dimensionality and complexity, achieves near-optimal performance.