Sean O'Connor

It seems there is a certain inefficiency in the use of weights in current neural networks. You are trying to come up with n linear classifiers all operating on one vector, if n is 200 or something you are going to find it very difficult to find that number of worthwhile different linear splits. Finding that number of nonlinear splits would seem more plausable. But then you have to use random projections https://discourse.processing.org/t/flaw-in-current-neural-networks/11512

I guess current artificial neural networks have inherited a lot of bagage from the 1950s onward. I think the foundations have just been taken on trust and never really looked at again or checked. Anyway: https://discourse.processing.org/t/flaw-in-current-neural-networks/11512