A dark room whose walls are covered with rows of hollow plaster faces lit from above, each of which appears to be a normal convex face
Unit 15 · Advanced + Cognitive Neuroscience · 65 min

The Predictive Brain

The brain as a prediction machine, from illusions to dopamine
Image: Pseudopanax, Wikimedia Commons, public domain

A hollow mask looks like a normal face, a pill without a drug eases pain, and a monkey's dopamine neurons fire at a light that announces juice rather than at the juice itself. This unit works through the idea behind all three: the brain predicts its own input and learns from the difference between prediction and fact. It computes a Bayesian perception by hand, reads the dopamine recordings of 1997, runs the temporal difference rule trial by trial, weighs the placebo and the hallucination experiments as tests of the idea, and ends with the free-energy principle and its critics.

Preview: the beginning of the unit

A hollow mask looks like an ordinary face from behind, a sugar pill eases pain, and the dopamine neurons of a monkey fire at the light that announces a drop of juice rather than at the juice itself. The three findings come from three corners of psychology, and in the last thirty years one idea has been used to connect them. The brain does not wait for its input and then react to it; it predicts the input and learns from the difference between the prediction and what arrives. This unit works through that idea with the numbers behind it, from Helmholtz’s inference to the arithmetic of the dopamine signal, and ends with the attempt to make it a theory of everything the brain does, together with what its critics say.

1 Perception as inference

Unit 3 ended with the conclusion that perception does not copy the image on the retina but infers the scene that most probably produced it, a view that Hermann von Helmholtz (1821-1894) called unconscious inference in 1867. The modern form of the idea treats the inference as a calculation with probabilities. The visual system holds expectations about the world that were learned before the current image arrived, it receives an image that fits some scenes better than others, and what we see is the scene that the two sources favour together. The rule for combining them is the one Unit 6 used for a diagnostic test, Bayes’ theorem, applied now to a scene instead of a disease.

The calculation uses three probabilities, and the rest of the unit refers to them by name. The prior is the probability of a scene before the current input is considered, which expresses what the system has learned about the world (faces bulge outwards, light comes from above, objects keep their size as they move). The likelihood is the probability of the input given a scene, which expresses what the eyes report and how reliable the report is. The posterior is the probability of the scene given the input, and the claim of the Bayesian view is that perception is the scene with the highest posterior.

The theorem says more than that expectations colour perception; it says by how much. A prior that covers a narrow range of possibilities pulls the posterior strongly, a likelihood that covers a narrow range pulls it the other way, and the side with the smaller spread wins. This weighting by reliability, which the field calls precision, is the part of the account that later sections test. The precision of a source is the inverse of its variance, which is the square of its spread (the standard deviation of Unit 1), and when the prior and the likelihood are both bell curves the posterior is itself a bell curve whose peak is the average of the two peaks weighted by their precisions. A prior with half the spread of the likelihood therefore has four times its precision, so the two peaks are averaged with weights of 4 and 1, and the posterior lies \(4/(4+1)\), which is four fifths, of the way towards the prior; two sources of equal spread meet halfway. The worked case runs the computation with the natural frequencies of Unit 6, that is, with counts out of 1,000 cases instead of probabilities.