The Predictive Mind By Jakob Hohwy |
The prediction error minimization idea says that all the brain ever does is minimize the error of predictions about its sensory input, formed on the basis of an internal model of the world and the body. The better these predictions are, the less error there is. On this view, the bottom-up signals in the brain, beginning with the sensory input, are conceived as prediction errors that work as feedback to the models maintained in the brain and their top-down predictions.
This is a simple idea, but with extreme explanatory ambition. Perception is the process of refining the models so that better predictions are formed, attention is predicting the sensory input precisely, and action is changing the input to fit with the model’s predictions. In this sense, perception, attention, and action are all best conceived in terms of statistical inference: the brain must use statistical tools to make sound inferences about the world based only on its sensory input.
To be meaningful, inference must be guided by reliable signals so a key part of inference is to estimate the precisions (or variance) of the prediction error (or sensory input). This implies the brain must keep track of the varying levels of noise and uncertainty in the environment, and adjust accordingly how much it relies on the sensory input according to such expected precisions. Mechanistically, this is tied to adjusting the gain on sensory input. Functionally, it is attention.
Simply put, the brain’s predictions of how precise the sensory input is control the “gates of perception”: we attend to signals that are expected to be precise. Optimizing expected precisions is difficult statistically because it is itself a type of inference, in need of further assessment of precision. To avoid regress, at some point, levels of such inference about inference must become uninformative. This induces a certain fragility to a prediction error minimizing system, which suggests that some psychopathologies, and some neurodevelopmental problems may be tied to inference on expected precisions.
Parts of the book pursue this idea, in particular for delusions and elements of autism. I speculate that some aspects of schizophrenia may be tied to low expectations of precision, or little trust in the sensory signal, and that some aspects of autism may be tied to too much trust in the sensory signal. This could potentially explain some of the symptoms of these disorders. A key element here is not chronic expectations of precision, but the ability to adjust the expectation for precision, depending on context. In our own lab, we explore these ideas, in particular concerning autism.
Jakob Hohwy |
Versions of all the linked papers are available from my website.
The book is available in paperback, hardback and kindle; it is also available from amazon.co.uk (amazon.de and amazon.com soon).
Jakob wrote: "The prediction error minimization idea says that all the brain ever does is minimize the error of predictions about its sensory input, formed on the basis of an internal model of the world". That idea is central to the 2-factor theory of delusion (see my post on this). If I suffer a stroke which disconnects my face recognition system from my autonomic nervous system, then a prediction error will occur next time I see my wife's face. So there must be something wrong with my internal model of the world. Hence that model needs to be updated so that it in the future it makes the correct prediction. Updating consists of adopting the belief that the person who looks like my wife is actually a stranger. That leads me to predict no autonomic response when I see someone with this face, and that prediction is correct i.e. prediction error has been minimised by adoption of this belief, even though the belief is delusional (Capgras delusion). Analogous prediction-error stories can be offered for other forms of monothematic delusion.
ReplyDelete"Brains compile, categorize and conjure."
ReplyDelete