Coherent ICA
In combining the Info max and Imax contributions to the objective function J(Wα, Wb), we bypassed the need for including regularization that provides a tradeoff between the Info max and Imax components we did so in order to simplify the formulation of the ICA algorithm. How could this objective function be modified to preserve the statistical dependence between the outputs of networks α and b, yet include regularization in the objective function? What are the implications of this extension?