On the parameter optimization in machine learning techniques: advances and paradigms

Abstract

Machine learning techniques have been actively employed in the last years in several research areas. Although such techniques are known for their good generalization skills, some of them are parameter-dependent, which may result in a high computational burden for learning such set of parameters, making them inviable for large-scale problems, as well as for real-time and prompt responses. This proposal aims at studying meta-heuristic optimization techniques, as well as their applications for machine learning optimization techniques, such as model and kernel selection, feature selection and information fusion, among others. This project also aims at investigating the suitability of optimizing parameters in the context of deep learning architectures, such as Convolutional Neural Networks, Restricted Boltzmann Machines, Deep Belief Networks and Auto-encoders. The idea is to focus on multi-objective optimization techniques, since just a few number of works have addressed such context. The proposal has the collaboration of Prof. Xin-She Yang from Middlesex University, England, which is a very active researcher on meta-heuristic-based optimization. The main idea is to validate his new algorithms in the context os this project. (AU)