Training Neural Networks is $\exists\mathbb R$-complete
This is a foundational result for theoretical computer science and machine learning, explaining why common techniques for NP-complete problems fail in neural network training.
The paper precisely determines the algorithmic complexity of training neural networks by proving it is ∃ℝ-complete, meaning it is equivalent to solving systems of polynomial equations and inequalities with real unknowns, and implies it is not in NP if ∃ℝ is larger than NP.
Given a neural network, training data, and a threshold, it was known that it is NP-hard to find weights for the neural network such that the total error is below the threshold. We determine the algorithmic complexity of this fundamental problem precisely, by showing that it is $\exists\mathbb R$-complete. This means that the problem is equivalent, up to polynomial-time reductions, to deciding whether a system of polynomial equations and inequalities with integer coefficients and real unknowns has a solution. If, as widely expected, $\exists\mathbb R$ is strictly larger than NP, our work implies that the problem of training neural networks is not even in NP. Neural networks are usually trained using some variation of backpropagation. The result of this paper offers an explanation why techniques commonly used to solve big instances of NP-complete problems seem not to be of use for this task. Examples of such techniques are SAT solvers, IP solvers, local search, dynamic programming, to name a few general ones.