Hybrid Neural Network Architectures for Image Recognition

Current convolutional networks work with inefficient pixel-wise image representation, which
does not provide almost any invariance. This leads to using very large training sets and
massive augmentation. We propose to decompose intra-class variances into two degradation
operators where one of them can be mathematically modelled by a superposition integral with a
transformation of the coordinates. We propose to design hybrid network architectures that use
both pixel-level and newly developed high-level invariant image representations such that the
high-level representation will eliminate the influence of modelable degradations. The other intraclass
variances will be tackled by deep learning on the pixel-level part of the network. We
suppose to develop multi-branch parallel architectures as well as single-branch ones, that we
obtain by generalization of group equivariant networks. This shall lead to a substantial reduction
of the training set without sacrificing the recognition rate. The results of the project could define
new standards in image-oriented network architectures.