We performed several simulations with feed-forward neural networks usi
ng an idealized tracking apparatus with tracks invariant under transla
tion and roto-translation transformations. Input information was provi
ded to the networks without any preprocessing. We implemented 2 and 3
layer architectures up to 50000 connections, and we tested the influen
ce of parameters such as learning rate, momentum, number of learning f
iles and noise rejection on the classification efficiency. The general
ization ability-is not so good as expected, whereas the classification
efficiency is larger than 90% for almost all the architectures, the i
nfluence of the above mentioned parameters being less than 10% overall
except for the noise rejection for which it increases up to 20%.