Analyzing perception and neural coding using adaptive experiments
By Christopher DiMattina (Florida Gulf Coast University)
Abstract: Quantitative approaches to brain and cognitive science often give rise to computationally intensive problems of model estimation and comparison. A growing body of literature has demonstrated that adaptive stimulus optimization methods hold tremendous promise for making these problems more tractable. In this talk, I give a brief overview of these techniques and some of their applications in sensory neuroscience and psychophysics, and discuss my recent research along these lines. Firstly, I present two recently developed novel implementations of the popular PSI algorithm for adaptive stimulus generation which make it tractable to performing adaptive psychophysical experiments using multi-dimensional stimuli. Secondly, I argue that given an accurate neural encoding model and perceptual decoding model, it is possible in principle to use psychophysical experiments to recover tuning properties of sensory neurons and to compare competing hypotheses about neural coding. I illustrate this with a specific example of characterizing the effects of contrast on orientation discrimination, and show that one can use psychophysical experiments to accurately recover the parameters of the contrast gain tuning functions which agree with neurophysiological data. I demonstrate that stimuli which are adaptively optimized for model comparison using an information-theoretic criterion are far more effective for distinguishing competing hypotheses of neural tuning than those presented using the method of constant stimuli. Finally, I point to interesting directions for future theoretical and experimental research in this growing field at the intersection of neuroscience, psychology and machine learning.