Corlett, Frith & Fletcher (2009) (henceforth CFF) expand on this idea and speculate on the biochemical substrates of each part of the process. They view perception as a “handshake” between top-down and bottom-up processing. Top-down models predict what we’re going to see, bottom-up models perceive the real world, then they meet in the middle and compare notes to calculate a prediction error. When the prediction error is low enough, it gets smoothed over into a consensus view of reality. When the prediction error is too high, it registers as salience/surprise, and we focus our attention on the stimulus involved to try to reconcile the models. If it turns out that bottom-up was right and top-down was wrong, then we adjust our priors (ie the models used by the top-down systems) and so learning occurs.He examines the possibilities that schizophrenia is a problem of increased bottom-up signalling with decreased top-down signalling, and autism a disorder of not tolerating even small differences between top-down and bottom-up's conclusions and refusing to shake hands.
They argue that autism is a form of aberrant precision. That is, confidence intervals are too low; bottom-up sense-data cannot handshake with top-down models unless they’re almost-exactly the same. Since they rarely are, top-down models lose their ability to “smooth over” bottom-up information. The world is full of random noise that fails to cohere into any more general plan.In the final section, Alexander sees some weaknesses in both possibilities.
If you like that sort of thing, I think this is his followup.