Forward and backward prediction in learning and perception

Predictive processing frameworks have emphasized the role of ‘forward’ prediction as a critical ingredient for learning and perceptual inference. Specifically, we anticipate sensory events that are likely to be presented in the future on the basis of the current sensory events. By comparing these forward predictions against incoming input, we can obtain an accurate estimate of the environment (i.e. perceive) and improve the predictions themselves (i.e. learn). Interestingly however, research in the field of statistical learning has taught us that backward predictive relationships - reflecting the probability of past events given present events - govern learning too. This work questions the privileged status of forward-looking mechanisms and renders it essential to consider the implications for our models of learning and perception. Here we discuss commonalities and differences between implications for learning and perception. We conclude that while forward and backward predictive relationships both shape learning, we retrieve future, but not past, predicted states during perception.