notes-cog-hechtNielsenConfabulationTheory

see http://www.scholarpedia.org/article/Confabulation_theory_%28computational_intelligence%29

This is a mostly belief net/belief maintanence system in which each variable can be assigned a number of mutually exclusive symbols (e.g. "the apple is colored ___" is a variable, and symbols might be "red", "green", and "yellow"), each symbol has a current activation representing degree of belief, and pairs of symbols from different variables have conditional probability between them (e.g. the 'red' symbol in the 'color' variable might have a link to the 'red delicious' symbol in the 'kind' variable), and results are discretized via a winner-take-all competition that selects the single most probable symbol for a given variable.

Like naive Bayes, incoming activations are simply summed, e.g. i think (not too sure here) that we effectively assume that each 'incoming' symbol is conditionally independent of all of the others give any choice of the symbol that we are trying to choose.

The wrinkle is that the winner-take-all competition (termed 'confabulation') can be run part-way, i.e., each variable has an associated 'current degree of confabulation' exogenous parameter which may be changed over time, and when this parameter is low, the activation associated with each symbol is just its probability, but when it is high, the winner-take-all-ness suppresses the less activated symbols in favor of the more activated ones. Since the different variables are connected, as the confabulation parameter is turned up for one variable, this reduces the effect of the suppressed symbols on the other variables. This means that, by turning up the confabulation parameter slowly over many variables at once (like simulated annealing), the system can converge to a consistent assignment to many variables.

Going beyond the belief net component of confabulation theory, the theory also postulates that symbols can be linked to 'actions', and also has some postulates on the neurobiological implementation of the system, for example, that the control of the confabulation parameter is a repurposing of the neurobiological systems for motor control, that an increase in the confabulation parameter corresponds to muscle contraction, that variables (called 'modules') correspond to thalamocortical modules, that symbols correspond to subsets of activated neurons within these modules, and that conditional probability links (called 'knowledge links') correspond to cortical synapses, managed in some way by an associated patch of thalamus.