Storage capacities and applications of some neural networks.

In this thesis, the storage capacities of the Bidirectional Associative Memories (BAM) and the Hopfield network and the applications of the multi-layer feed-forward network are studied and presented. There are four sub-topics in this thesis, the first one is the use of ring and cascade architectures in storing temporal or ordered patterns. A comparison of these architectures is presented. The second sub-topic is to use some new methods to increase the storage capacities of the BAM and the Hopfield network. The uses of a modified Hebb rule and multi-threshold values are studied. Comparisons among the Hebb rule, the delta rule and these new methods are also presented. Results show that these new methods can store ten 35-pixel images while the Hebb rule cannot. The third sub-topic is the use of a multi-layer feed-forward network with the back-propagation training to do pulse compression. Results of this approach are presented. Lastly, the use of a multi-layer feed-forward network to do filtering is also studied and the results are presented. Source: Masters Abstracts International, Volume: 30-03, page: 0840. Supervisor: H. K. Kwan. Thesis (M.A.Sc.)--University of Windsor (Canada), 1990