Wilson S. Geisler¹ (), Anqi Zhang¹; ¹University of Texas at Austin

Quantitative models of perceptual performance often include limitations due to the physical properties of the stimuli, limitations due to sensory processing, and the assumption of rational (Bayes optimal) decision processes applied to the outputs of the sensory processing. Determining the Bayes optimal decision process is important for characterizing the available information, how that information varies with stimulus conditions, and what specific computations achieve optimal task performance. Although there are tasks where humans reach near optimal performance there are many where they fall short. How is it that humans achieve near optimal performance in some cases and yet clearly fail in other cases? One approach to this question is Bayesian heuristic decision analysis (BHDA), which systematically explores the space of heuristic decision processes, using Bayes optimal decision processes as the benchmark. If there are biologically plausible heuristics that approach Bayes-optimal performance, then they provide realistic testable hypotheses for how humans can achieve near optimal performance. Conversely, if there are no such biologically plausible heuristics, then the expectation is non-optimal performance. We demonstrate this analysis on visual search tasks. In covert search with a single known target, the optimal decision rule is to weight the feature responses at each potential target location by the discriminability (d’) at that location, add the log of the prior probability at that location, and then pick the location with the maximum value. We find that a wide range of simple decision heuristics closely approach optimal accuracy, even though these heuristics largely ignore the actual variation in discriminability and prior probability. These results help explain the high efficiency of humans in this task. Bayesian heuristic decision analysis also shows that efficient fixation selection in overt search can be achieved with realistic limitations in memory and posterior updating. Several other important implications of this approach will be described.

Acknowledgements: Supported by NIH grants EY11747 and EY024662