Guessing games of the brain..
These are some reading notes of the article 'Whatever next? Predictive brains, situated agents, and the future of cognitive science' by Andy clark
Some of the text below may be directly extracted from the article.
The whole function of the brain is summed up in: error correction - W Ross Ashby
Abstract Brains are prediction machines. Incoming sensory input are matched with top down predictions by a hierarchical generative model to minimize the prediction error within a bidirectional cascade of cortical processing. This forms a unifying model of perception and action. The article critically examines the 'hierarchical prediction machine'.
In mamalian brains, prediction error seems to be corrected by a cascade of cortical processing where higher levels attempt to predict the lower level inputs, on the basis of their own emergent models of the causal structure of the world (signal source).
Each element of the neural cascade creates a model on the basis of the signals received so far. With this model, each element predicts the next input and the error between the predicted and recieved signal is used to update the model. A large collection of units predicting and correcting itself yields a rich body of information about the source of the signal (i.e. the world)
A model like this follows Helmholtz(1860), by depicting perception as a process of probabilistic, knowledge driven inference.
Key idea of Helmholtz - sensory systems infer the causes from their 'bodily effects'. In this paradigm, brain does not build its world model by accumulating bottom up low-level cues such as edge-maps, but by high to low level mapping. -Hohwy 2007 (Chater & Manning 2006, p. 340)
This idea of Helmholtz was pursued further by the advances in machine learning - back propagation (McClelland et al. 1986; Rumelhart et al. 1986) and on to Helmholtz machine (Dayan et al. 1995; Dayan & Hinton 1996; see also Hinton & Zemel 1994)
Helmholtz machine -> learn new representations in a multi-level system without a large labelled dataset. It sought to improve upon back prop.
It had its own top-down connections to provide the desired states for the hidden units, in effect self supervising the development of its perceptual recognition model using a generative model that tried to create the (fantasy) sensory pattern for itself.
A generative model aims to capture the statistical structure of some set of observed inputs by tracking the causal matrix responsible for that very structure.
The Paper (http://goo.gl/lxFc9d)
Andy Clark Perceiving as Predicting (https://youtu.be/05P41FQlgjI?t=11m58s).