Theory of Neural Networks
Abstract
A new neural unsupervised learning technique has been proposed in this work. Technique is based on the hierarchical partition of the patterns. Each partition corresponds to one neuron, which is in general a higher-order neuron. The partition is performed by iterating the neuron weights in an attempt to maximize a defined criterion function. The method is implemented on several examples and is found to give good results. In the second implemented example the method obtained a good solution, whereas the traditional adaptive resonance method and self-organizing maps produced unsatisfactory results. The method is fast, as it takes typically from about 2 to 5 iterations to coverage. Although the proposed method is prone to get stuck in local minima, this did happen in the simulations in only very difficult problems and this problem could be solved by using gradient algorithms for searching for the global maximum, like the Tunneling Algorithm.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jul 31, 1991
- Accession Number
- ADA253187
Entities
People
- Amir F. Atiya
- Yaser S. Abu-mostafa
Organizations
- California Institute of Technology