Random Variable Analogy Measurement Based ICA

In this paper we propose the analogy measurement of two random variables, and discuss the principle of maximizing nongaussianity of observed data to estimate independent components sequentially based on unsupervised learning neural network. We also prove the non-polynomial moment theorem in a generalized scheme, and reveal the feasibility that replaces the analogy measurement by the expectation of a non-quadratic smooth even function G(·) based on the theorem. A formula to compute the sign of above algorithm is given, which overcomes the contradiction between the objective function and the sign computation formula.