The number of cases depends on number of descriptors. It is evaluated that for neural networks, there should be 10 times more cases than connections between neurons (so the number of cases depends on the network achitecture too). My calculations with the use of ANN has given quite good accuracy of predictions having twice more cases than connections. For simple cases (not so much predictors, and few neurons only) 100-200 cases it is sufficient dataset. But if you'd like to feed the net with files (not with numbers; i.e. deep learning) several thousand of cases (files) are necessary.

Regards

Hubert