CNNs have produced great results. But a problem they have is that they require too much data. Could there be ways to initialize their weights such that this requirement is loosed? This code checks some ideas, for example: can we initialize the CNN using the weights of the CAE (as it was already proposed in the original CAE article)? Do we actually get better results if we do that? And what about Greedy layer-wise pre- training? But how to adapt it to the convolutional nature of CNNs?

[still under construction........]