Three steps towards reliable and efficient systems for machine learning algorithms

Resumen

This thesis compiles the work realised during four years of research in machine learning algorithms and techniques. The aim of this study lays in how to execute the inference process of machine learning algorithms in a more efficient way and how to achieve more reliable techniques. For that, this thesis consist in three main steps: first, the design and implementation on an FPGA of an accelerator that takes advantage of the optimisation opportunities offered by sparsity in DNNs, second, the design and implementation on an FPGA of an accelerator for gradient boosting decision trees in the context of hyperspectral images classification, and third, an analysis of bayesian networks for hyperspectral images classification to demonstrate how the uncertainty metrics can help us in many tasks to achieve more reliable networks.