Abstract
We present a neural net architecture that is organized as a hierarchical cascade of layers of simple threshold units. A node at layer k + 1 is connected to many nodes at layer k and thus asserts the existence of a subpattern at layer k. It is a well-known technique to cascade such layers in a compositional hierarchy so that the highest layer nodes assert the presence of complete patterns composed of many subpatterns. In our implementation, feedback links from higher to lower layers and intralevel inhibitory connections between competing subpatterns are included. If the net is driven by an appropriate iterative update scheme, it acts as an associative memory with the stored patterns as stable states. Such a structure is well suited for an optical implementation. It may be possible to decompose a given recognition task that requires space-variant connection patterns into a cascade of layers connected by space-invariant patterns. The space invariance allows 2-D node planes to be easily connected optically to other 2-D node planes. We propose an optical realization and discuss problems in the implementation of the threshold operation, the connection patterns, and the handling of negative values. Results of some simulations are presented.
© 1986 Optical Society of America
PDF ArticleMore Like This
Demetri Psaltis, Jeffrey Yu, Xiang Guang Gu, and Hyuk Lee
TuA3 Optical Computing (IP) 1987
David Casasent
MJJ1 OSA Annual Meeting (FIO) 1990
Xiang Y. Yang, Francis T. S. Yu, and Don A. Gregory
TuK3 OSA Annual Meeting (FIO) 1991