Neural Network Simulation at Warp Speed: How We Got 17 Million Connections per Second
Abstract
We describe a fast back-propagation algorithm for a linear array of processors. Results of an implementation of this algorithm on Warp, a ten processor, programmable systolic array computer, are reviewed and compared with back-propagation implementations on other machines. Our current Warp simulator is about 8 times faster at simulating the NETtalk text-to-speech network than the fastest back-propagation simulator previously reported in the literature. This fast simulator on Warp is being used routinely in a road recognition experiment for robot navigation at Carnegie Mellon. Our results indicate that linear systolic array machines can be efficient neural network simulators. Planned extensions and improvements to our current algorithm is discussed. (EG)
Document Details
- Document Type
- Technical Report
- Publication Date
- Mar 31, 1988
- Accession Number
- ADA218906
Entities
People
- David S. Touretzky
- Dean A. Pomerleau
- George L. Gusciora
- H. T. Kung
Organizations
- Carnegie Mellon University