On the Study of Semantics in Deep Learning Models

Abstract

Deep learning architectures have been extensively studied in the last years, mainly due to their discriminative power and effectiveness in many crucial problems in computer vision, such as face and people identification, as well as object recognition, just to name a few. However, one problem related to these models concerns with their number of parameters, which can easily reach thousands of hundreds. Another drawback is related to the need for large datasets for training purposes, as well as their high probability of overtraining, mainly because of their complex architecture. Although some recent works have proposed different solutions to alleviate this problem, such approaches still require parameters, which need to be fine-tuned, and also are skilled-dependent. This proposal aims at "learning how these techniques do learn", i.e., to learn their good points and shortcomings. Techniques such as Restricted Boltzmann Machines and Deep Belief Nets will be studied in order to obtain a better understanding of their working mechanism through meta-heuristic-based optimization. This proposal also comprises an internship at Middlesex University, United Kingdom.