Context Dependent Pattern Recognition - A Framework for Hybrid Architectures Bridging Chaotic Neural Networks Based on Recursive Processing Elements and Symbolic Information

摘要

This work discusses a hybrid structure that conjugates connectionist associative memories and deterministic automata, for the implementation of context dependent pattern recognition. The associative component of the hybrid system is built through coupled recursive maps with bifurcation and chaotic dynamics (Recursive Processing Elements - RPEs). Its output feeds a deterministic state machine that controls the context of the pattern recognition tasks and produces related symbolic outputs. The proposal is illustrated in a scenario for context dependent (visual) pattern recognition, performed by an autonomous agent. Such "Learner" agent alternates between contexts of unsupervised image recognition and contexts of interaction with a "Teacher" agent, in supervised sections of image recognition. Computational experiments and related measures show the effectiveness of the proposal.