Behaviour of feed-forward neural networks in invariant track finding

Abstract We performed several simulations with feed-forward neural networks using an idealized tracking apparatus with tracks invariant under translation and roto-translation transformations. Input information was provided to the networks without any processing. We implemented 2 and 3 layer architectures up to 50 000 connections, and we tested the influence of parameters such as learning rate, momentum, number of learning files and noise rejection on the classification efficiency. The generalization ability is not so good as expected, whereas the classification efficiency is larger than 90% for almost all the architectures, the influence of the above mentioned parameters being less than 10% overall except for the noise rejection for which it increases up to 20%.