Abstract
If the spatial validity of prior information is varied systematically, does human behavioral performance improve in a graded fashion, and if so, does visual cortex represent the probability directly? Cortical activity was measured with fMRI while subjects performed a contrast-discrimination task in which the spatial validity of a prior cue for target location was systematically varied. Subjects viewed four sinusoidal gratings (randomized contrasts of 12.5, 25, and 50%) shown in discrete visual quadrants presented twice. The contrast in one location (target) was incremented in one of the two presentations. Subjects reported with a button press which presentation contained the greater contrast. The target grating was signaled in advance by a cue which varied in spatial validity; at trial onset, small lines pointed to four, two, or one of the possible target locations, thus indicating the target with 25, 50, or 100% probability. Behavioral performance was 2.1 and 3.3 times better in the 100% probability condition than the 50% and 25%, respectively (p < .001, ANOVA). Unlike behavioral performance, cortical activity in early visual areas showed the same increase in response amplitude for cued versus uncued stimuli for both 100% and 50% probability (V1-V4, V3A all p < .18, Student's t-test, 25% had no uncued condition). How could behavioral performance improve in a graded fashion if cortical activity showed the same effect for different probabilities? A model of efficient selection in which V1 responses were pooled according to their magnitude rather than as a simple average explained the observations (AIC difference = −15).