This fit was obtained using standard MLP back-propagation with learning rate and momentum adjustments. This method doesn't converge to a best solution and takes a very long time to train. The points near the center require many hyperplanes to fit and standard back-prop falls down.