An acoustooptic technique was used to design and construct a prototype of a processor capable of multiplying a 128 x 128 complex-element matrix by a 128 complex-element vector. The device was specified to execute 10(5) matrix-vector products per second with 8-bit resolution. The performance of each component of the prototype was evaluated, as was the impact of component performance on system performance.

2SECURIT'e C A5FiPCAT ON AuL7HORIWY 3 D'5PiL,-i1 avN A~a A 2b DECLASSiFCA71ON D>OWdNGRA01NG SCHED LE Aoproved for public release; distribution unlimited. 4 PERFORMING ORGANIZATION REPORT' NUMBER(S) 5 VION17ORING DRGAN ,ZA-SN NRL Memorandum Report 6372 6a NAME OF PERFORMING ORGANIZATION 60 09F< CSYVMBOL 'a NAME 3F 'AO"J'OP 'C Z r2 (If applicable) Naval Research Laboratory Code 6330 6c. ADDRESS (City, State. and ZIP Code) Th AOODPESS City, State :in, zR C.;0e' Washington, DC 20375-5000 8a. NAME OF FUNDING, SPONSORING 8b OFF CE S*MB0L 9 ' 0 OC-REMENT7 NS-RLEN-:)E,-CA-OC% ORGANIZATION O (f applicable) Office ofNavalTechnologv _______ 8C. ADDRESS (City, State, and ZIP Code) 0SOuRCE OF 'N '4V~NBERS POOGRA IPOEC-A ~ .và rigoV 21-00:EEM EN 14 NO NO 'OASO NO Arigon A221-00PE 62314 RJ 14C37 DN380-06 2 7 1 I ~iTLE (include Security Classficatnon)

Higher-order neural networks1have several desirable computational characteristics such as significantly decreased convergence times, increased storage capacity over first-order layered networks (which respond only to weighted sums of the inputs ∑ W i x i ), and an ability to encode a prioriknowledge in the network. Simulation results for the problem of learning to classify clusters in a binary string showed that a second-order network (which also includes weighted sums of products ∑ W i j x i x j ) generalized correctly 90% of the time after training on only 10% of the training set.