Parametrised Neural Network Design and Compilation into Hardware
Most artificial neural networks consist of one or more arrays of components, each of which is obtained by replicating a few simple processing elements connected together in a uniform manner. This paper illustrates the use of Ruby, a language of relations and functions, for describing such networks and for implementing them in hardware. Our objective is to enable designs to be rapidly realised and evaluated.
KeywordsParallel Composition Sequential Circuit Simple Processing Element Repeated Composition Neural Network Hardware
Unable to display preview. Download preview PDF.
- Algotronix Limited, CAL 1024 Datasheet, 1990.Google Scholar
- Baji, T. and Inouchi, H., “Systolic Processor Elements for a Neural Network”, US Patent 5,091,864, 25 February 1992.Google Scholar
- Boser, B.E., Sackinger, E., Bromley, J., IeCun, Y. and Jackel, L.D., “Hardware Requirements for Neural Network Pattern Classifiers”, IEEE Micro, February Issue, pp. 32–40, 1992.Google Scholar
- Holt, J.L. and Hwang, J.N., “Finite Precision Error Analysis of Neural Network Hardware”, to appear in IEEE Trans. Neural Networks, 1992.Google Scholar
- Jones, G. and Sheeran, M., “Circuit Design in Ruby”, in Formal Methods for VLSI Design, J. Staunstrup (ed), North-Holland, pp. 13–70, 1990.Google Scholar
- Luk, W., “Systematic Serialisation of Array-Based Architectures”, to appear in Integration, Special Issue on Algorithms and VLSI Architectures, 1992.Google Scholar
- Luk, W. and Page, L., “Parametrising Designs for FPGAs”, in FPGAs, W. Moore and W. Luk (ed), Abingdon EE&CS Books, pp. 284–295, 1991.Google Scholar
- Murray, A.F., Smith, A.V.W. and Butler, Z.F., “Bit-Serial Neural Networks”, Proc.BIPSConf., pp. 573–583, 1987.Google Scholar
- Skubiszewski, M., “A Hardware Emulator for Binary Neural Networks”, Proc. FPL 92, Vienna, 1992.Google Scholar