Codes and Expansions (CodEx) Seminar

Wojciech Czaja (University of Maryland)
Fourier Scattering as Efficient Feature Extraction

We present a construction of a family of feature extractors which combine Mallat's scattering transform framework with the benefits of different time-frequency representations. We do this by introducing a class of frames, called uniform covering frames, which includes a variety of semi-discrete Gabor systems and other Fourier-based representations. We then incorporate these frames into an iterative neural network-like structure, to generate our candidate features, which we aggregate into a new scattering transformation. This approach proves advantageous in several data-related applications, in particular in the context of hyperspectral imaging and similar imaging modalities. We also explore several other mathematically-inspired ideas, including those of composite wavelets and rotationally invariant Fourier frames. This talk presents a body of joint work with several co-authors: Weilin Li, Ilya Kavalerov, Mike Pekala, and Rama Chellappa.