The neural network loading problem is undecidable
Herbert Wiklicky · 1994
There exist several well known results concerning the computational complexity of training, design, etc. of (artificial) neural networks which demonstrated that these problems are at least NP complete. Extending these results we will prove here that the training of neural networks, i.e. the so called "loading problem", in general, is even "unsolvable". This implies in particular that there exists no general or universal training algorithm which for any given neural network architecture could determine a correct set of "weights" such that a certain desired input/output behavior of the neural network is achieved. 1 Introduction Artificial neural networks, like the Multi Layer Perceptron (MLP), can be used for approximating any function belonging to certain "reasonable" classes (e.g. continuous or measurable functions) as closely as desired. Related results can be found for example in [8] etc. These results are often celebrated as the theoretical justifications of what lead to a renaissa...