Fine tuning deep boltzmann machines through meta-heuristic approaches

Abstract

The Deep learning framework has been widely used in different applications from medicine to engineering. However, there is a lack of works that manage to deal with the issue of hyperparameter fine-tuning, since machine learning techniques often require a considerable human effort in this task. In this paper, we propose to fine-tune Deep Boltzmann Machines using meta-heuristic techniques, which do not require the computation of the gradient of the fitness function, that may be insurmountable in high-dimensional optimization spaces. We demonstrate the validity of the proposed approach against Deep Belief Networks concerning binary image reconstruction.