Abstract
A non-linear neural network model that employs Gaussian-type threshold function is presented. The model is characterized by a probability-natured dataflow in the computing units of the network. It has the ability of learning from unlabeled input signals for pattern classification and functional association. The network consists of two neuron layers: (1) a Gaussian functional net which generates an internal representation of input patterns, and (2) a binary winner-take-all net which provides deterministic output of the network. A significant test approach is applied to the self-organization process of the network. The unsupervised learning scheme employs a bi-variate optimization technique. It tries to minimize the entropy of the network and a two-way complementary criteria function. The network is especially advantageous to the classification of noise-corrupted and incompletely represented stochastic patterns.
Original language | English (US) |
---|---|
Pages | 515-519 |
Number of pages | 5 |
State | Published - 1990 |
Externally published | Yes |
Event | Proceedings of the Twenty-First Annual Pittsburgh Conference Part 4 (of 5) - Pittsburgh, PA, USA Duration: May 3 1990 → May 4 1990 |
Other
Other | Proceedings of the Twenty-First Annual Pittsburgh Conference Part 4 (of 5) |
---|---|
City | Pittsburgh, PA, USA |
Period | 5/3/90 → 5/4/90 |
ASJC Scopus subject areas
- General Engineering