On the existence of infinitely many realization functions of non-global local minima in the training of artificial neural networks with ReLU activation

Abstract: Gradient descent (GD) type optimization schemes are the standard instruments to train fully connected feedforward artificial neural networks (ANNs) with rectified linear unit (ReLU) activation and can be considered as temporal discretizations of solutions of gradient flow (GF) differential equations. It has recently been proved that the risk of every bounded GF trajectory converges in the training of ANNs with one hidden layer and ReLU activation to the risk of a critical point. Taking this into account it is one of the key research issues in the mathematical convergence analysis of GF trajectories and GD type optimization schemes, respectively, to study sufficient and necessary conditions for critical points of the risk function and, thereby, to obtain an understanding about the appearance of critical points in dependence of the problem parameters such as the target function. In the first main result of this work we prove in the training of ANNs with one hidden layer and ReLU activation that for every $ a, b \in \mathbb{R} $ with $ a < b $ and every arbitrarily large $ \delta > 0 $ we have that there exists a Lipschitz continuous target function $ f \colon [a,b] \to \mathbb{R} $ such that for every number $ H > 1 $ of neurons on the hidden layer we have that the risk function has uncountably many different realization functions of non-global local minimum points whose risks are strictly larger than the sum of the risk of the global minimum points and the arbitrarily large $ \delta $. In the second main result of this work we show in the training of ANNs with one hidden layer and ReLU activation in the special situation where there is only one neuron on the hidden layer and where the target function is continuous and piecewise polynomial that there exist at most finitely many different realization functions of critical points.