A Library for Fast Kernel Expansions with Applications to Computer Vision and Deep Learning. Part 1a.

The main contribution of the thesis is the development of a fast library for approximating kernel expansions, which enables the use of Kernel Methods in large-scale datasets. Kernel Methods are computational costly for big data, this library enables the use of non-linear features in log-linear time. This approximation is based on the Walsh Hadamard. A SIMD implementation of the Fast Walsh Hadamard that outperforms current state-of-the-art methods has been developed. The thesis contains interesting applications to Computer Vision and Deep Learning which can serve as guideline for novel researchers in statistical learning.