Perception is a process of inference, whereby incoming sensory evidence is interpreted based on prior expectations about the sensory world. Thus, the neural code of perception should be evaluated based on how optimally it computes perceptual inference. However, the neural code of perception has conventionally been evaluated by its capacity to represent sensory information faithfully. Due to this misalignment in the computational goal of perception, assessments of the neural code have been biased towards discriminability over generalizability, and efficiency over robustness. In this review, I suggest ways in which we can evaluate the neural code based on its capacity to achieve the goal of perception, that is, perceptual inference.
H. Shin (Wed,) studied this question.