Abstract
Human ability to detect a target among distractors is influenced by several well-investigated factors, including set size, target-distractor similarity, and distractor heterogeneity. A factor that has been ignored is the uncertainty of individual items in a single display. Intuitively, items that provide less reliable visual information should, on any given trial, be assigned less weight in the target detection judgment. In earlier work, we showed that a Bayesian model of search makes this intuition precise and accurately predicts human performance in homogeneous displays containing items of differing reliability (i.e., all distractors have the same orientation, but vary in contrast). Here, we test the Bayesian model when distractors are heterogeneous and drawn from a near-uniform feature distribution. We again find that humans integrate information across space nearly optimally. The standard MAX model from signal detection theory allows for neither differing reliabilities nor heterogeneous distractors.
Furthermore, we propose a neural implementation of Bayesian visual search using probabilistic population codes. In this framework, each item elicits activity in a population of neurons with so-called Poisson-like variability. On each trial, an entire probability distribution over the stimulus is automatically encoded by each population pattern. Since the Bayesian computations are very complex, approximations are needed for a neural network to implement them. We consider networks with one of three types of operations: 1) linear; 2) quadratic; 3) quadratic plus divisive normalization. We impose the constraint that the output is again in the Poisson-like format, to facilitate downstream computation. We find that the third type outperforms the first two, as measured by percent information loss with respect to the Bayesian observer. This is true for both homogeneous distractors and heterogeneous distractors drawn from a uniform distribution. Together, these results show that Bayesian theories of perception have great potential to be extended to complex integration tasks.