Training Fully Connected Neural Networks is existsR-Complete

We consider the algorithmic problem of finding the optimal weights and biases for a two-layer fully connected neural network to fit a given set of data points. This problem is known as empirical risk minimization in the machine learning community. We show that the problem is ∃R-complete. This complexity class can be defined as the set of algorithmic problems that are polynomial-time equivalent to finding real roots of a polynomial with integer coefficients. Furthermore, we show that arbitrary algebraic numbers are required as weights to be able to train some instances to optimality, even if all data points are rational. Our results hold even if the following restrictions are all added simultaneously.
∙ There are exactly two output neurons.
∙ There are exactly two input neurons.
∙ The data has only 13 different labels.
∙ The number of hidden neurons is a constant fraction of the number of data points.
∙ The target training error is zero.
∙ The ReLU activation function is used.
This shows that even very simple networks are difficult to train. The result explains why typical methods for NP-complete problems, like mixed-integer programming or SAT-solving, cannot train neural networks to global optimality, unless NP=∃R. We strengthen a recent result by Abrahamsen, Kleist and Miltzow [NeurIPS 2021].