Abstract
Sensory information is noisy and so inferences about the visual world benefit from repeated sampling and integration of evidence into a cumulative ‘decision variable’ or DV. Computational modelling and single-unit recording studies show that the DV scales with mean signal strength. However, for many psychophysical stimuli (e.g. random dot kinematograms) this quantity is confounded with signal variability, a quantity important for estimating the reliability of visual information. To assess how the DV is computed during visual categorisation, we asked human observers to discriminate multi-element arrays in which the mean and variance of a decision-relevant dimension (shape or colour) were manipulated orthogonally. Observers were slower to respond to more variable arrays, and weighted outliers less heavily in the decision. This replicates recent modelling work, which showed that observers accumulated the total log posterior ratio (LPR) associated with the two alternatives, calculated from the sigmoidal likelihood function linking elements to choices, rather than accumulating mean evidence or signal-to-noise ratio. Here, we additionally used functional magnetic resonance imaging (fMRI) to assess the validity of the LPR model at the neural level. Analysis focused on the ventromedial prefrontal cortex (vmPFC), a brain region where blood-oxygen level dependent (BOLD) signals correlate with the probability of receipt of positive feedback or reward. BOLD in the vmPFC correlated with the mean LPR associated with the array, even when the raw mean and variance of both decision-relevant and decision-irrelevant dimensions were included as nuisance covariates. This suggests that observers do indeed calculate the log posterior ratio during categorical choice. However, we also observed dissociable BOLD correlates of the signal mean and variance in the parietal cortex, suggesting that the statistics of the array are computed there. Together, these data provide neural evidence that humans make categorical judgements in a statistically optimal fashion.
Meeting abstract presented at VSS 2012