Visual decision making in the presence of stimulus and measurement correlations

Our brains process sensory information to infer the state of the world. However, the input from our senses is noisy, which may lead to errors in perceptual judgements. A number of theoretical studies have modeled perception as a process of probabilistic inference that involves making decisions based on uncertain evidence. Bayesian optimality is a general principle of probabilistic inference that has been successfully used to build quantitative models of perception. In addition, several experimental studies show that human observers make best possible decisions, and hence exhibit close to Bayes-optimal behavior on various visual perceptual tasks such as visual search, sameness judgement, and change detection. However, the impact of structured stimuli on decision-making remains largely unexplored. Moreover, the sensory measurements can themselves be strongly correlated to produce a structured representation of the stimulus input. These measurement correlations can interact with the structure of the external input in many possible ways and should not be considered in isolation.

In this work, we focus on visual search task to examine how visual perception is affected by structured input. We analyze the responses of subjects on a target detection experiment where the stimulus orientations were generated with varying strength of correlations across different experimental sessions. We fit several models to the experimental data using maximum-likelihood parameter estimation. We use rigorous model selection to find that human observers take into account stimulus correlations in detecting a target. However, they behave suboptimally in inferring the correct stimulus correlations that were used in the experiment. We find that perhaps observers treat the partial stimulus correlations identically and behave differently when the stimuli are perfectly correlated.

We also describe how the relation between measurement and stimulus correlations affects the performance of an ideal Bayesian observer in a family of target detection tasks. We find that the effect of measurement correlations depends on its interaction with stimulus correlations and other statistical structure parameters. Measurement correlations always improves the performance of the ideal observer on a detection task with multiple targets; whereas in the case of single target, the impact is significant only in the presence of strong external structure.