Application of New Artificial Neural System Information Processing Principles to Pattern Classification.
Abstract
Pattern self organization (without a teacher) and temporal compression methods using aritificial neural networks have been demonstrated on Lenglish (a form of artifical speech), based on the text from a childrens book. This report demonstrates how neural networks can self organize on non-syncronized time patterns, in effect, how to stabilize spatiotemperal patterns in time, which then permits adaptive networks to lean. The work utilized the TRW enhanced Mark III-1 neurocomputer to host and run the network simulations. A novel transform based on differential pjase was developed to be time shift invariant, but which retained local phase relationships. The results of this work can be applied to dimensionality reduction, bandwidth reduction, to self-generating encoding/decoding. Iy is felt that a network of this type could be used to perform Reed-Solomon decoding.
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 20, 1987
- Accession Number
- ADA180442
Entities
People
- Michael Myers
- Robert Kuczewski
- William J. Crawford