[54] NEURAL NETWORK FOR VOICE AND PATTERN RECOGNITION
[75] Inventors: Masakatsu Maruyama; Hiroyuki Nakahira; Masaru Fukuda; Shiro Sakiyama, all ol Osaka, Japan
[73] Assignee: Matsushita Electric Industrial Co.,
Ltd., Osaka, Japan
[21] Appl. No.: 08/864,938
[22] Filed: May 29, 1997
[30] Foreign Application Priority Data
May 30, 1996 [JP] Japan 8-136822
[51] Int. C I. G06E 1/00
[52] U.S. CI 706/25; 706/15; 706/25;
706/27
[58] Field of Search 706/15, 25, 27,
706/41, 42
[56] References Cited
U.S. PATENT DOCUMENTS
4,479,241 10/1984 Buckley 382/159
4,774,677 9/1988 Buckley 706/23
4,989,256 1/1991 Buckley 706/41
5,161,203 11/1992 Buckley 382/157
5,204,938 4/1993 Skapura et al 706/42
5,325,464 6/1994 Pechanek et al 706/41
5,375,250 12/1994 Van Den Heuvel 706/41
5,487,153 1/1996 Hammerstrom et al 395/670
5,517,600 5/1996 Shimokawa 706/15
5,524,175 6/1996 Sato et al 706/41
5,608,844 3/1997 Gobert 706/42
5,649,069 7/1997 Gobert 706/41
5,748,849 5/1998 Gobert 706/27
FOREIGN PATENT DOCUMENTS
5-61847 3/1993 Japan .
OTHER PUBLICATIONS
Korn, Granino A., "Neural Networks and Fuzzy-Logic Control on Personal Computers and Workstations", The MIT Press, Cambridge, Massachusetts, pp. 70, Jan. 1995. Hrycej, Tomas, "Modular Learning in Neural Networks: A Modular Approach to Neural Network Classification", John Wiley and Sons, Inc., pp. 36, Jan. 1992.
Primary Examiner—Tariq R. Hafiz
Assistant Examiner—Wilbert L. Starks, Jr.
Attorney, Agent, or Firm—McDermott, Will & Emery
[57] ABSTRACT
A neural network circuit for performing a processing ol recognizing voices, images and the like comprises a weight memory for holding a lot ol weight values (initial weight values) which correspond to a plurality ol input terminals ol each ol a plurality ol neurons forming a neural network and have been initially learned, and a difference value memory for storing difference values between the weight values ol the weight memory and additionally learned weight values. The weight memory is formed by a ROM. The difference value memory is formed by a SRAM, for example. During operation ol recognizing input data, the initial weight values ol the weight memory and the difference values ol the difference value memory are added together. The added weight values are used to calculate an output value ol each neuron ol an output layer. Accordingly, the initial weight values can be additionally learned at a high speed by existence ol the difference value memory having a small capacity. Thus, new numerals, characters and the like can be recognized well without error.
15 Claims, 18 Drawing Sheets
