Quantum Neurodynamics: A New Neural Network Paradigm for Knowledge Representation, Cognition, and Control. Final Report on Phase 1.

摘要

The objective of the project is to design a computable solution to a very large and very important class of adaptive estimation and control problems which lie at the foundation of machine intelligence and robotics. Researchers describe a new neural network architecture called the Parametric Avalanche (PA) which is designed to implement a constructive innovations method solution to the general stochastic filtering problem for nonlinear systems in arbitrary noise. The problem of representing multitudes of state evolution equations within the network is solved by incorporating one very robust and well-understood evolution equation-the Nonlinear Schrodinger (NLS) equation--into the network dynamics, and using associative memory methods to encode the observed system's dynamics into its trajectories. By coupling the prediction error into the scalar potential field of the NLS, the continuing observation deflects the recalled trajectory toward the minimum (locally) mean squared error estimator for the observed process, within the capacity constraints of the architecture. The architecture may be incorporated into an adaptive control module which is capable at least of balancing an inverted pendulum. Using networks of these PA modules, explanatory models for the function and dysfunction of biological cognitive systems are constructed.