Low Computational Cost Machine Learning: Random Projections and Polynomial Kernels

Abstract

[EN] According to recent reports, over the course of 2018, the volume of data generated, captured and replicated globally was 33 Zettabytes (ZB), and it is expected to reach 175 ZB by the year 2025. Managing this impressive increase in the volume and variety of data represents a great challenge, but also provides organizations with a precious opportunity to support their decision-making processes with insights and knowledge extracted from massive collections of data and to automate tasks leading to important savings. In this context, the field of machine learning has attracted a notable level of attention, and recent breakthroughs in the area have enabled the creation of predictive models of unprecedented accuracy. However, with the emergence of new computational paradigms, the field is now faced with the challenge of creating more efficient models, capable of running on low computational power environments while maintaining a high level of accuracy. This thesis focuses on the design and evaluation of new algorithms for the generation of useful data representations, with special attention to the scalability and efficiency of the proposed solutions. In particular, the proposed methods make an intensive use of randomization in order to map data samples to the feature spaces of polynomial kernels and then condensate the useful information present in those feature spaces into a more compact representation. The resulting algorithmic designs are easy to implement and require little computational power to run. As a consequence, they are perfectly suited for applications in environments where computational resources are scarce and data needs to be analyzed with little delay. The two major contributions of this thesis are: (1) we present and evaluate efficient and data-independent algorithms that perform Random Projections from the feature spaces of polynomial kernels of different degrees and (2) we demonstrate how these techniques can be used to accelerate machine learning tasks where polynomial interaction features are used, focusing particularly on bilinear models in deep learning.