This notebook demonstrates the need for multi-layer ANN when the solution involves fitting many non-linear edges.

I created a Jupyter notebook for using an image to monitor the progress of an ANN to solve a complex classification. It is available here or viewed online at GitHub.

The images typically show a vertical structure, suggesting that the 2 inputs are not treated equally. It could be that this is due a random initialization: in one case there has a horizontal characteristic.

The images demostrate that a single hidden layer has difficulty handling the non-linear edges of the classification. The addition of additional layers significantly improves the handling of these edge cases. It should be noted that percentage-wise most sample points are handled well by the SHL. But the proper handling of the edge cases may be an important in certain problems.