logo
surprise_minimisation_theory

The Dark Room

One of the most interesting theories of brain function in recent years is that of surprise minimisation. This theory was born of computational neuroscience which brings together aspects of physics, information theory and machine learning and tells us that your brain spends its time predicting what will happen next and then updating predictions based on what your senses tell it. In every situation your brain is testing and updating models of reality so that surprises are kept to a minimum. These models are based on the probabilities of different outcomes and are thought to be Bayesian. This just means that instead of assigning probabilities based on the frequency of events you start with some assumptions and continuously update based on new evidence.

This is all well and good but it has led some people to wonder why, if this model is correct, do we not seek out a quiet dark room where nothing happens and surprises of all kinds are kept to a minimum? This is known as the Dark-Room Problem.

I recently came across an unusual paper (Friston et al 2012) on this subject in the format of a 3-way conversation between a physicist: Karl Friston, an information theorist: Christopher Thornton and a philosopher: Andy Clark. The very simple answer to this paradox is that surprise is relative to what you expected in the first place, if you did not realise you were in a dark room this in itself would be surprising. All life has evolved within a specific econiche, as Friston calls them, and therefore carries around a model or rather is a model of that environment. Your predictions can only be based on what is normal for you. This paper is worth the effort as it very nicely demonstrates what Neuroscience is today and how it comfortably lives at the cross-roads of Information Theory, Physics and Philosophy.

Friston has been a major contributor to the field in developing the theory of free-energy minimisation (Friston 2010) in a Bayesian brain (to give it its full name) which he was led to through early work in schizophrenia. This is interesting as schizophrenia can be seen as   prediction error leading to hallucinations and difficulty in attributing agency to sensory inputs including your own voice. Understanding this may open the door to new treatments and indeed to early diagnosis using alterations in, for example, EEG (Riera 2012).