The hypothesis that the brain is a Bayesian inference machine has been around for a long time. When I hear some of the people leading the AI research, they often refer to the brain as an statistical engine.

Main ideas

  • There is growing evidence that the brain represents inputs probabilistically; animals, humans included, exist in a world of sensory uncertainty. Our brains must deal with uncertainty all the time, and they generate representations of the world around us. These representations guide our actions.
  • The idea behind the Bayesian approach is that information provided by sensory inputs from the outside world is represented by a conditional probability density function over the set of unknown variables - the posterion density function.
  • We receive inputs from several external sources over space and time: vision, sound, touch, smell, etc. Each input is statistically independent, but they can be combined into a posterior density function which is proportional to the product of multiple functions.
  • Not every source is equally reliable; we are naturally biased towards the most reliable sources.
  • Vision: a brain perceives a 2D image and based on two-dimensional patterns of light and infers depth based on the relative brightness of the light sources.
  • Our brains “commits” to a representation of reality (side note: I’ve met my share of people are committed to their own strange versions of reality. The stranger the reality, the more committed they seem to be).

Main paper

Further reading

  • Friston, K. (2010). The free-energy principle: a unified brain theory? Nature Reviews Neuroscience. The foundational paper for active inference and FEP.
  • Hohwy, J. (2013). The Predictive Mind. Oxford University Press. A book-length treatment of predictive processing and its implications for philosophy of mind.
  • Rao, R.P.N. & Ballard, D.H. (1999). Predictive coding in the visual cortex: a functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience. One of the earliest concrete proposals for predictive coding in neural circuits.