In the current iteration, this only implements feed forward neural networks -- which are basic and can only resolve basic correlations.

For more advanced stuff (like image or video processing) you have to move to recurrent neural networks.