random data as test data for a NN

Can I use random generated numbers to test a neural network? If I have a small number of samples (around 400), this data probably has a bias and therefore the NN should learn the relationship between input and output?

Thanks