Why Minds Break · Ian Pitchford · Chapter 4 of 10

The Society of Mind, Revisited

Name: Why Minds Break
Author: Ian Pitchford

The chapter asks how a mind built from simpler parts is organised, separates the claims hidden in words like module and hierarchy, and tests predictive accounts of the brain.

Marvin Minsky's rule for explaining a mind is that no part may itself be intelligent, or the theory hides a person inside, like the chess automaton Edgar Allan Poe suspected. The chapter needs such an account because the developmental changes of Chapter 3 reach experience and behaviour only through whatever does the perceiving and acting. It traces the idea of a module from Jerry Fodor's list of features, through Peter Carruthers' weaker massive modularity and Barrett and Kurzban's emphasis on functional specialisation, to Michael Anderson's evidence that brain regions are reused across tasks. Specialisation, it concludes, need not mean a sealed, inborn mechanism.

Comparative anatomy sets limits but is easily over-read. How large the human brain is for its body depends on which scaling model is chosen, and studies measuring prefrontal volume and prefrontal neuron numbers disagree in a dispute the chapter calls unsettled. Larger primate brains have sparser long-range connections. Cortical columns and von Economo neurons look distinctive but say little about function on their own. Studies by Hasson and by Murray show that brain areas integrate information over different spans of time, without implying one clock per level.

The central proposal is to take a process, an operation turning specified inputs into outputs, as the unit of explanation and to describe each candidate along six separate axes: evolutionary history, developmental timing and plasticity, learning rate, temporal integration, anatomical organisation and computational depth. None can be read off another. A worked example, judging the reliability of someone who sometimes keeps promises, shows how a model with a volatility level must beat a simpler learner with a fixed learning rate. Reading, acquired culturally yet fast, is one counterexample to any single ladder.

John Hughlings Jackson's hierarchy of re-representation and Daniel Dennett's tower of Darwinian, Skinnerian, Popperian and Gregorian creatures are read as orderings serving different purposes, not as anatomical floors. The predictive account, running from Helmholtz through Rao and Ballard to Karl Friston's free energy principle, is broken into prediction, Bayesian inference, predictive coding and active inference, which need separate evidence. Following Del Giudice, accurate prediction is not equated with fitness.

The evidence review finds expectation effects well documented but the proposed architecture of separate prediction and error neurons largely undemonstrated, with reviewers still divided. The chapter treats the free energy principle as a modelling framework whose process theories bear the empirical risk, while reporting critics who reject that reading. Domain-general control is compatible with Minsky's rule, and a module claim is defined so that it can fail. Predictive models thus receive the ordinary burden of competing with rivals and being allowed to lose, and their entries in the research registry of Chapter 10 are graded claim by claim.

Questions the chapter asks

What does it mean to call part of the mind a module, and how could such a claim be shown false?
Does the large human prefrontal cortex explain distinctively human abilities?
Is the predictive brain theory supported by evidence, or can it be made to fit anything?

This is a summary of the chapter, written for this site; it is not an extract from the book. Coming soon from Amazon in Kindle, paperback and hardback editions.