I yeaaaarn for a future where we will have a general parallelizable training method for neural networks (or even better a principled way to initialize trained w

I yeaaaarn for a future where we will have a general parallelizable training method for neural networks (or even better a principled way to initialize trained weights like the work being done on wavelet scattering). Long training times with backpropagation is a serious obstacle when doing experiments. I had hoped DFA would be it but it doesn't work for image tasks sadly.