Hybrid Optical Inference Machines
Abstract
This program has investigated the use of limit cycles to represent and processing symbolic information in the context of an inference machine. This approach was proposed as a means of overcoming problems with fault tolerance and relatively small space-bandwidth products in current spatial light modulator (SLM) technology. The program has focused on developing a storage medium with many limit cycles (oscillatory modes) available and a method for coupling the various modes in a desired way. Because of their flexibility, neural network ideas were used as the basis for the components and algorithms developed. In the theoretical realm, the program has had many accomplishments. First, the self- oscillating neural network (SONN) model was developed and characterized as the oscillatory medium. This model was designed with optical spatial SLMs in mind and does not require any training or programming. Furthermore, it is highly tolerant of static parameter variations inherent in the optics. Next, the spectral back-propagation (SBP) training algorithm was developed with complete generality as a means of forming the coupling trajectories. This algorithm trains input-output sequences into a network using an error criterion based on a Fourier series decomposition of the sequences. The method allows the interconnects to have trainable time delays in addition to the weights.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 27, 1991
- Accession Number
- ADA243985
Entities
People
- Cardinal Ware
- James Kottas
- Vernon Shrauger
Organizations
- Massachusetts Institute of Technology