Attentional Priority as Expected Information Gain

We have long known that neurons in the parietal and frontal eye fields provide representations of visual priority, which sparsely encode the locations of stimuli worthy of attention or gaze. However, the computational definition of priority is entirely unclear. Terms like “behavioral relevance” and “salience” that are invoked for explaining priority are only given ad hoc definitions that often conflict across tasks. I will propose a computationally grounded definition of priority as the expected information gain (EIG) of an information channel in a particular context. I will describe evidence that neurons in the lateral intraparietal area encode the two ingredients of Bayesian EIG, prior uncertainty and stimulus diagnosticity, and do so independently of expected rewards. I will end by presenting a neurocomputational model of how an EIG-based priority representation may emerge from interactions between the fronto-parietal and executive networks.