Figure 1.

The framework of Deep IDA. Classes are represented by shapes and views are represented by colors. The (DNN are used to learn nonlinear transformations of the D views, the outputs of the DNN for the views ($f^d$) are used as inputs in the optimization problem, and we learn linear projections ${\mathbf{A}}_d,d=1,\dots ,D$ that maximally correlate the nonlinearly transformed views and separate the classes within each view.