摘要：We demonstrate that a neural network composed of only three nodes and three connections arranged in a 2-inputs, 1-middle, 1-output architecture is able to perform differentiation of univariate functions (Đ). Using a proposed empirical technique, we assess the network’s generalization capability by approximating a functional form for the growth function ΔF(N). We calculated the probability of error to be ≈10^-7 allowing us to justify the effectiveness of the simplistic approach in modeling a non-trivial task such as Đ.
关键词：architecture;differentiation;generalization;growth function
