Energy-based learning models and their applications

Abstract

Energy-based models have been employed in a number of research areas, since they belong to a broad framework that comprises both probabilistic and non-probabilistic techniques. Among the most used techniques from this context, one can refer to the Restricted Boltzmann Machines and their deep learning-driven variants: Deep Belief Nets and Deep Boltzmann Machines. Such models also have shortcomings that will be considered in this proposal, such as the automatic optimization of their parameters, how to enhance the effectiveness of the learning process by means of sampling in Markov chains, as well as how to obtain a better regularization process based on different temperatures and pruning neurons. The proposal has two main lines of research: (i) applications and (ii) theoretic work. In the former, we will look for applications that have not evaluated the aforementioned energy-based models to date, mainly the ones related to biometric- and medicine-oriented data. In regard to the theoretic research, we expect to contribute with different models for sampling in Markov chains aiming at enhancing the inference step of these models. The proposal also considers national and foreign researchers, as well as undergraduate and graduate students. Since there are a very few research groups in Brazil that work in the same context of this proposal, we believe we can make difference and help the scientific community with works related to energy-based learning. (AU)