Learning from error predictions rather than prediction errors: a theory of self-supervised learning in the brain

Humans learn by doing and observing, but also by imagining the result of simulated actions in their mind. When learning from executed or observed actions, traditional theories posit that behavior improves by minimizing the difference between (i) the perceived outcome derived from sensory feedback and (ii) the desired outcome (performance error) or the outcome predicted by an internal model (prediction error). Supporting these theories, decades of electrophysiological studies have documented the role of the dopaminergic system and cerebellum in generating activity patterns compatible with the hypothesized error signals. In comparison, our understanding of the neural basis underlying the ability to learn from simulated actions is far more limited. Here, I argue that this form of self-supervised learning is best described in terms of a quantity known as the predicted performance error (PPE), theorized more than 30 years ago and yet largely overlooked ever since. Complementary to prediction errors (PEs), PPEs reflect “error predictions” between predicted and desired outcomes. Unlike PEs, however, PPEs exist even in the absence of sensory feedback and can in principle serve to evaluate actions without executing them, based on an efference copy of the underlying motor plan. I review evidence supporting the computation of PPEs in the brain, and propose a network implementation based on a cortico-basal ganglia loop interacting at multiple timescales with prefrontal, parietal, and hippocampal areas. I then connect PPEs to recent discoveries on motor planning, and formulate novel predictions linking PPEs to various high-level brain functions such as metacognition, problem-solving and deception.