Abstract
Theories of neural information processing generally assume that sensory input is processed along hierarchical stages that start with analog representations and gradually transition to task-related, abstract representations. While the neural code of such abstract information remains unclear, neurophysiological findings suggest that a scalar code could be used to encode behavioral relevance. Here we test this hypothesis in human fMRI studies, using data from five feature-based attention tasks where participants selected one feature from a compound stimulus containing two features. We found that the majority of voxels in a cortical area showed consistently higher response when subjects attended one feature over the other. We examined this biased coding across brain areas, participants, and stimulus domains, and found robust bias within brain areas, consistent direction of such bias across areas for a given participant, and similar bias for multiple feature types (e.g. color, motion directions and objects). Using a receiver operating characteristics analysis to quantify the magnitude of the bias, we found stronger bias in frontoparietal areas than in visual areas, indicating more abstract representations in high-level areas. We also examined the contribution of this bias to multivariate decoding by removing the mean response from each condition before applying pattern classification. We observed decreased classification accuracies in frontal and parietal areas, but not in visual areas. Our results suggest a gradient coding mechanism along visual hierarchy, where high-dimensional coding in sensory cortex allows fine-grained representation of feature attributes, and low-dimensional (possibly one-dimensional scalar) coding in association cortex facilitates the simple read-out of decision and control variables.
Acknowledgement: R01EY022727