A Computer-Assisted Formulation of an Abstract Model for Some Aspects of Neocortical Learning

Abstract

This dissertation presents an abstract model for some aspects of neocortical operation. Contributions in three areas have been attempted: Knowledge representation, learning mechanisms, and neocortical modeling. Recognizer/ Predictors (RPs), data structures resembling small, uniform frames, but applied in parallel, are proposed to represent knowledge. Pattern induction is proposed as a mechanism for learning and correcting RPs, by detecting and abstracting regularities from episodic memories of RP usage. RPs and pattern induction are described in the context of MUL, a computer program modeling infant-like development in a simple environment. The abstract concepts of RPs and pattern induction are used to suggest a new model for neurons of the neocortex. In this model, neocortical neurons begin in an unfixed state in which they have no meaningful output. Each neuron records episodes, storing the activity the neuron receives at its input synapses from fixed neurons. Each neuron performs pattern induction upon its collection of episodes until it discovers some suitable regularity; it then undergoes fixation, by modifying its input synapses to recognize (become activated by) future instances of that regularity. Fixed neurons correspond to RPs. Among other mechanisms introduced are default activation and inhibition. Default activation of an RP (or fixed neuron) involves activation in the absence of an instance of the regularity recognized by that RP (or fixed neuron). Such default activation is shown to be useful in initiating motor responses, approximation, generalization, and anticipation of future experience. Inhibition is a mechanism to correct improper defaulting behavior; inhibition is implemented by inhibitory RPs (or inhibitory neurons), and the learning of inhibitions is also accomplished by a pattern induction process. Among the advantages claimed for the proposed neocortical model: -A definite, unchanging meaning is assigned to each neuron. -Neurons may discover regularities in the activities of other neocortical neurons, to recognize progressively more complex concepts. -Neural activation patterns may be meaningfully recorded as episodic memories. -Pattern induction provides a plausible mechanism to assign meanings to neurons, even in the absence of direct feedback or obvious examples, provided by an external "teacher", and in spite of noise or malfunction. -Pattern induction, and the resulting RPs (or fixed neurons), are useful independently of the meaning of the inputs to the RP or neuron. -The proposed neocortical model is compatible with biochemical and neurophysiological evidence of neocortical function.