Implementing Artificial Neural Networks in Integrated Circuitry: A design Proposal for Back-Propagation
Abstract
In an attempt to develop CMOS circuitry (both analog and digital) for the implementation of artificial neural networks, the back-propagation learning algorithm was examined in detail. Simulations were performed to determine to robustness of this algorithm to anticipated implementation artifacts such as quantization and weight-range limitation. Circuitry which computes with analog signals and digitally encoded weights was then designed to implement the algorithm within the tolerances determined by the simulations. The architecture of an alternative, fully digital design was also defined and its performance compared with that of both the analog/digital design and a fully analog design based on circuitry that has been proposed in the literature.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 18, 1988
- Accession Number
- ADA202541
Entities
People
- S. L. Gilbert
Organizations
- Massachusetts Institute of Technology