Abstract
Perceptual interpretations of the environment emerge from the concerted activity of neural populations in decision-making areas downstream of sensory cortex. When the sensory input is ambiguous, perceptual interpretations can be biased by prior beliefs that reflect knowledge of environmental regularities. These effects are examples of Bayesian reasoning, an inference method in which prior knowledge is leveraged to optimize decisions. However, it is not known how decision-making circuits combine sensory signals and prior beliefs to form a perceptual decision. To address this, we studied neural population activity in the prefrontal cortex of two macaque monkeys trained to report perceptual judgments of ambiguous visual stimuli under different prior statistics. Monkeys judged whether a visual stimulus was oriented clockwise or counterclockwise from vertical and communicated their decision with a saccadic eye movement towards one of two visual targets. The meaning of each response option was signaled by the target's orientation (clockwise vs counterclockwise) and was unrelated to its spatial position. Because the spatial configuration of the choice targets varied randomly from trial to trial, changes in prior stimulus statistics biased the animals' perceptual reports, but not the overt motor responses. We analyzed the component of the neural population response that specifically represents the formation of the perceptual decision (the decision variable, DV), and found that its dynamical evolution reflects the integration of sensory signals and prior beliefs. The DV’s initial value before stimulus onset reflects the prior belief in the future state of the sensory environment, while the dynamic range of the DV's ensuing excursion reflects the relative influence of the incoming sensory signals. These findings reveal how prefrontal circuits integrate prior stimulus expectations and incoming sensory signals at the behaviorally relevant timescale of the single trial, thus exposing a general mechanism by which prefrontal circuits can execute Bayesian inference.