Computational limitations on training sigmoid neural networks (Q685499)

We investigate the computational limitations of training small neural networks with sigmoid neurons. We focus on 2-layer networks with 3 neurons which interpolate a set of labeled training vectors where each label is in \(]0,1[\). In the presence of additional parameter constraints (e.g. demanding that all parameters are binary numbers) we show that we cannot efficiently find parameters for the network such that the network would produce labels consistent to the given set of training vectors (unless \(P=NP\)). Thus, we perform an analyses which comes to quite similar results as in the area of threshold networks.