We prove a theorem concerning the approximation of multivariate continuous functions by deep ReLU networks, for which the curse of the dimensionality is lessened. Our theorem is based on the Kolmogorov--Arnold superposition theorem, and on the approximation of the inner and outer functions that appear in the superposition by very deep ReLU networks.