Boolean Structured Principle Components Analysis (BsPCA)

Layer by layer training has been very popular since Jeffrey Hinton time. Layerby layer training enables us to visualize and understand how a deep learningnetwork works.To develop robust and optimal layer by layer training, we have to ensurethat the training is convex and it preserves the distances in the manifold of thedata. We train each layer using Principal Components Analysis.Our BsPCA model inherits from BSnet model family that maximizes trainingresults by ensuring the convexity of the overall training loss.