Universal Approximation Property of Random Neural Networks

In this paper, we study random neural networks which are single-hidden-layer feedforward neural networks whose weights and biases are randomly initialized. After this random initialization, only the linear readout needs to be trained, which can be performed efficiently, e.g., by the least squares method. By viewing random neural networks as Banach space-valued random variables, we prove their universal approximation properties within suitable Bochner spaces. Hereby, the corresponding Banach space can be more general than the space of continuous functions over a compact subset of a Euclidean space, namely, e.g., an $L^p$-space or a Sobolev space, where the latter includes the approximation of the derivatives. Moreover, we derive some approximation rates and develop an explicit algorithm to learn a deterministic function by a random neural network. In addition, we provide a full error analysis and study when random neural networks overcome the curse of dimensionality in the sense that the training costs scale at most polynomially in the input and output dimension. Furthermore, we show in two numerical examples the empirical advantages of random neural networks compared to fully trained deterministic neural networks.