1Wellcome Trust Centre for Neuroimaging, University College London
2Gatsby Computational Neuroscience Unit, University College London
3Institute of Opthalmology, University College London
In Bayesian models of perceptual decisions, a belief distribution over the stimulus is combined with the value of each outcome to yield a choice that maximises expected value. Psychophysical data demonstrates that value is integrated with an estimate of internal uncertainty to optimally guide perceptual decisions (Whiteley & Sahani, 2008). How this integration is implemented by the human brain is unknown.
We measured fMRI BOLD responses while inducing shifts in perceptual decision criteria through asymmetric value. We used a blend of house and face images in order to generate a full psychometric function between objects with known stimulus-selective activation profiles in inferior temporal cortex. Subjects were not informed of the image continuum, and were simply asked to categorise stimuli as either faces or houses. Asymmetric monetary losses were imposed for incorrect face and house decisions, varying on a trial-by-trial basis. This design allowed us to ask how value and uncertainty interact to alter people’s perceptual decisions, and how these variables are encoded and integrated in the brain.
Behavioural results indicated shifts in the point of subjective equality as a result of asymmetric costs, as predicted by a Bayesian integration of value with sensory uncertainty. Functional imaging showed stimulus-selective activations in bilateral FFA and PPA that correlated with choice probability for faces and houses, respectively. However, activity in these regions was invariant to increases in stimulus-specific value. Instead, both types of asymmetric value trials (face and house) revealed consistent activations in left inferior frontal sulcus, anterior thalamus and posterior parietal cortex (PPC). In contrast, a stimulus uncertainty regressor derived from psychophysical data showed positive correlations with activity in anterior insula and paracingulate, consistent with previous reports (Grinband et al., 2006). A preliminary analysis suggested that the interaction of uncertainty and value is expressed in a modulation of activity in PPC.
Together, these data point to a theoretical framework in which value and uncertainty are integrated within dissociable prefrontal and cortico-thalamic networks at a late stage of a perceptual decision hierarchy. This post-sensory integration of value in perceptual decisions may be adaptively advantageous, allowing action selection to be flexible in its usage of sensory evidence in different reward contexts.