EVALUATION OF NETWORK ARCHITECTURES ON TEST LEARNING TASKS.

Two experiments with the gradient back propagation (GBP) algorithm have been carried out to determine how the architecture affects the performances of the network. The two examples have been designed to investigate two different properties of the networks: their memorization capacities and their ability to generalize by synthesizing appropriate predicates. The first experiment is an extension to the GBP algorithm of previous work comparing the memorization and generalization abilities of various network models on simple associative memory tasks. In the second experiment a network is taught to detect the presence of a given pattern in a signal.