Behavior and Learning in Networks with Differing Amounts of Structure.

摘要

This research is investigating how well large networks that are built from neuron-like elements can be made to perform, and learn to perform, by giving them different types and amounts of built in structure, and the ability to learn by generating new nodes in additions to changing weights. Substantial improvements in both learning speed and performance have been achieved on both pattern recognition problems and a range of problems typically used to demonstrate the power of connectionist networks. In addition, a number of new micro-circuits and sub-networks have been specified with which more powerful and more flexible networks can be built. These include: Back-cycling nets that handle learning(along with many useful functions), rather than have that handled by the system that executes the net; Nets that handle symbols as well as numbers; and Micro-circuits for productions and perceptual transforms.