Source: Medium When I led a DNN project based on few data ~20K examples for biometric ID, based on poor quality modality, I was trying to find out ways of incorporating prior knowledge to CNN in order to ease the data scale requirements for training.

Then I’ve found the ScatterNet idea which is a nice trick in terms of using a fixed and known structure of Wavelets based complex tree borrowed from the image processing domain.