Tag
This paper presents an exponential depth hierarchy for ReLU networks in terms of L2 approximation error, demonstrating that deeper networks offer exponentially improved representational power for function approximation.