Low power digital neuron for SOM implementations

  • Authors:

  • Roberta Cambio;David C. Hendry

  • Affiliations:

  • School of Engineering and Physical Sciences, University of Aberdeen, Kings College, Aberdeen, Scotland;School of Engineering and Physical Sciences, University of Aberdeen, Kings College, Aberdeen, Scotland

  • Venue:
  • ICANN/ICONIP'03 Proceedings of the 2003 joint international conference on Artificial neural networks and neural information processing
  • Year:
  • 2003

Quantified Score

Hi-index 0.01

Visualization

Abstract

As applications of the Self-Organising Map emerge in portable devices, power dissipation becomes a crucial design issue. The digital implementation of the SOM which is introduced in this paper meets low power requirements by means of increasing silicon area while reducing the number of clock cycles required to process each element of an input vector. Designs of a single neuron requiring two clock cycles, one clock cycle, and 1/2 clock cycle per element of the input vector are presented. The designs offer a reduction in power of a factor of 3 for an increase in silicon area of some 33%. The contribution of each routine composing training and classification to total power is also illustrated.