Learning gradients via an early stopping gradient descent method

We propose an early stopping algorithm for learning gradients. The motivation is to choose "useful" or "relevant" variables by a ranking method according to norms of partial derivatives in some function spaces. In the algorithm, we used an early stopping technique, instead of the classical Tikhonov regularization, to avoid over-fitting. After stating dimension-dependent learning rates valid for any dimension of the input space, we present a novel error bound when the dimension is large. Our novelty is the independence of power index of the learning rates on the dimension of the input space.