NONLINEAR PREPROCESSING OF INPUTS TO LINEAR NEURAL NETS,
Abstract
Discrimination of analog signal patterns by linear single-gain layer nets can be significantly improved by feeding the analog signals into preprocessors that convert each analog signal to a binary signal with m bits, thus increasing the number of gain elements from n to nm. The number of input vectors to which an arbitrarily desired net output can be assigned increases correspondingly from n to mn. This result holds for any binary converter with word length m independent of the code of the converter. The number increases further with the radix q if a q-ary preprocessor is employed. The common quantizer shows particular merits for practical applications as only one output line in any quantizer is active, thus allowing gain adjustments independent of each other within each quantizer. The combination of quantizer and linear net is reported on in detail. With a forced learning-type training algorithm, final gain values are shown to represent the difference of the conditional probabilities of the input pattern classes. The combination of quantizer and linear net instruments a type of likelihood ratio. With an error-correcting training algorithm, the final gains form a gain vector such that the error with respect to the desired output for each class becomes minimized in the least mean square sense. (Author)
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 1966
- Accession Number
- AD0645499
Entities
People
- F. D. Powell
- Johannes G. Goerner
- L. A. Gerhardt
Organizations
- Bell Aircraft Corporation