@ChrisDiehl They patched together two layers of SVMs and did some min-maxing over the representation in the hidden layer.