Using PHiPAC to speed error back-propagation learning

Author

Bilmes, Jeff ; Asanovic, Krste ; Chin, Chee-whye ; Demmel, Jim

Author_Institution

Dept. of Electr. Eng. & Comput. Sci., California Univ., Berkeley, CA, USA

Volume

5

fYear

1997

fDate

21-24 Apr 1997

Firstpage

4153

Abstract

We introduce PHiPAC, a coding methodology for developing portable high-performance numerical libraries in ANSI C. Using this methodology, we have developed code for optimized matrix multiply routines. These routines can achieve over 90% of peak performance on a variety of current workstations, and are often faster than vendor-supplied optimized libraries. We then describe the bunch-mode back-propagation algorithm and how it can use the PHiPAC derived matrix multiply routines. Using a set of plots, we investigate the tradeoffs between bunch size, convergence rate, and training speed using a standard speech recognition data set and show how use of the PHiPAC routines can lead to a significantly faster back-propagation learning algorithm

Keywords

backpropagation; convergence of numerical methods; digital arithmetic; matrix multiplication; neural nets; software libraries; software portability; speech recognition; ANSI C; PHiPAC; bunch size; bunch-mode backpropagation algorithm; coding method; convergence rate; error backpropagation learning; optimized matrix multiply routines; performance; portable high-performance numerical libraries; standard speech recognition data set; training speed; workstations; Computer errors; Convergence; Guidelines; Libraries; Production; Registers; Signal processing algorithms; Space exploration; Speech recognition; Workstations;

fLanguage

English

Publisher

ieee

Conference_Titel

Acoustics, Speech, and Signal Processing, 1997. ICASSP-97., 1997 IEEE International Conference on

Conference_Location

Munich

ISSN

1520-6149

Print_ISBN

0-8186-7919-0

Type

conf

DOI

10.1109/ICASSP.1997.604861

Filename

604861