[underaxon]

I may be wrong but I think this is what kernel methods (eg. SVM) do, right? So this looks like a (deep)SVM where the kernels are small NNs.