Abstract
Traditional accounts of selective attention hold that top-down mechanisms increase the gain of sensory neurons tuned to behaviorally relevant stimuli in order to facilitate perception. Recent theoretical and psychophysics studies suggest that selective attention operates more flexibly by targeting the most informative neurons, which – depending on task demands – are not always the most responsive (e.g., Navalpakkam & Itti, 2007; Scolari & Serences, 2009). For example, neurons tuned away from the target (off-channel neurons) are particularly informative when performing a difficult discrimination between two similarly oriented gratings. Furthermore, we previously showed that relatively higher off-channel activation in V1 predicted performance on a difficult discrimination task. However, given that attention was not manipulated here, two distinct models may account for these results. First, top-down attention might target off-channel neurons, increasing the amount of information available to downstream decision mechanisms (early gain account). Second, downstream decision mechanisms may favorably weight sensory input from informative off-channel neurons during decision-making in the absence of attentionally-modulated firing rate changes in early visual cortex (optimal read-out account). The current study tested these competing hypotheses using fMRI and a forward ‘encoding model’ to measure feature-selective modulations in early visual areas. Subjects performed either a difficult orientation discrimination task, contrast discrimination task, or rapid serial visual presentation letter task at fixation. Directing attention to the difficult orientation task resulted in a multiplicative scaling of feature-tuning functions in V1-V3v such that the activation of neural populations tuned to the target was enhanced. Contrary to our predictions, no evidence of top-down off-channel gain was observed. However, trial-by-trial fluctuations in the activation level of off-channel responses predicted perceptual decisions. Together, these results support the optimal read-out account: top-down attention inflexibly targets the most active sensory neurons, but downstream decision-related mechanisms dynamically read-out information from the most informative neurons based on task demands.