Learning Capability of a Simple Neural Network

Dranreb Earl Juanico, Christopher Monterola, Caesar Saloma


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.


architecture; differentiation; generalization; growth function

Full Text: