Reinforcing learning in Deep Belief Networks through nature-inspired optimization

Mateus Roder, Leandro Aparecido Passos, Gustavo Henrique de Rosa, Victor Hugo C. de Albuquerque, João Paulo Papa

Abstract: Highlights • Novel DBN with weight-based residual connections between layers. • Reinforcement and regularization of the information flow. • Application of metaheuristic optimization to fine-tune Res-DBN hyperparameters; Abstract Deep learning techniques usually face drawbacks related to the vanishing gradient problem, i.e., the gradient becomes gradually weaker when propagating from one layer to another until it finally vanishes away and no longer helps in the learning process. Works have addressed this problem by introducing residual connections, thus assisting gradient propagation. However, such a subject of study has been poorly considered for Deep Belief Networks. In this paper, we propose a weighted layer-wise information reinforcement approach concerning Deep Belief Networks. Moreover, we also introduce metaheuristic optimization to select proper weight connections that improve the network’s learning capabilities. Experiments conducted over public datasets corroborate the effectiveness of the proposed approach in image classification tasks.