Supervised-unsupervised combined neural learning for independent component analysis
Yang Chen, Zhenya He · 2004
A neural network approach to independent component analysis (ICA) is proposed. The supervised-learning backpropagation rule is used to train multilayer perceptron for approximating the signal distribution adaptively, giving an appropriate estimate of the nonlinear activation function in the unsupervised learning rule. A comparison with purely unsupervised learning is also made.