speculative model of dreams and hallucinations based on analysis by synthesis, arguing that the elaboration of content in these phenomena could be due to nothing more than a “disordered or random or arbitrary round of confirmation and disconfirmation” (p. 12). This has been spectacularly supported and refined by recent work at Google Research (e.g., Mordvintsev, Olah, and Tyka 2015). Today I could simplify my account: in a Bayesian network silence counts as confirmation. Whatever the higher levels guess counts as reality by default in the absence of disconfirmation.
