Modular Ring Architecture for Realization of Large Scale Neural Networks

Period of Performance: 01/01/1991 - 01/01/1993

$267K

Phase 2 SBIR

Recipient Firm

Caelum Research Corporation
1700 Research BoulevardSuite 250
Rockville, MD 20850
Principal Investigator

Research Topics

Abstract

PROGRESS IN REALIZATION OF LARGE SCALE, EXTENSIBLE NEURAL NETWORKS HAS BEEN FOILED BY PROBLEMS OF PROVIDING FOR MASSIVE INTERCONNECTIVITY AND FOR OVERCOMING FAN IN/OUT LIMITATIONS IN AREAEFFICIENT VLSI/WSI IMPLEMENTATIONS. THE MODULAR NEURAL RING (MNR) ARCHITECTURE, WHICH WAS PROPOSED AND INITALLY INVESTIGATED IN PHASE I OF THIS PROJECT, SOLVES THE CONNECTIVITY PROBLEMS WHILE PROVIDING FOR HIGH PARALLELISM, HIGH UTILIZATION OF HARDWARE AND MODULAR CONSTRUCTION OF LARGE SCALE ANNS. ITS REAL TIME DYNAMIC EXTENSIBILITY AND RECONFIGURABILITY ALLOWS FOR ECONOMICAL REALIZATION OF NEURAL MODELS AND MECHANISMS, IN APPLICATION AREAS SUCH AS VISION, NATURAL LANGUAGE PROCESSING AND DECISION MAKING. THE STRUCTURAL AND OPERATIONAL PROPERTIES OF THE MODULAR NEURAL RING ARCHITECTURE WERE INVESTIGATED DURING PHASE I OF THIS PROJECT. THE RESULTS OF PHASE I ESTABLISH CONVINCINGLY THE FEASIBILITY AND PRACTICALITY OF THE MNR ARCHITECTURE, ESPECIALLY IN COMPARISON WITH OTHER EXISTING IMPLEMENTATION ARCHITECTURES. THE OVERALL OBJECTIVE OF PHASE II IS TO RESEARCH, DEVELOP, DESIGN AND TEST THE ESSENTIAL HARDWARE AND SOFTWARE COMPONENTS OF THE MNR ARCHITECTURE FOR COST-EFFECTIVE REALIZATIONS OF LARGE SCALE ARTIFICIAL NEURAL NETWORK MODELS. COMPONENT AND SYSTEM PROTOTYPES WILL BE DEVELOPED, DESIGNED, TESTED AND EVALUATED. A NEW INTEGRATED CIRCUIT CHIP SET WILL BE DEFINED AND SPECIFIED FOR COMMERCIALIZATION OF MODULAR AND EXTENSIBLE IMPLEMENTATION OF LARGE SCALE NEURAL NETWORKS.