Effective fine-tuning training of deep Boltzmann machine based on spatial Monte Carlo integration

抄録

A deep Boltzmann machine (DBM) is a probabilistic deep learning model; DBM learning consists pretraining and fine-tuning stages. This study focuses on the fine-tuning stage, and it develops a new and effective fine-tuning method based on spatial Monte Carlo integration (SMCI), which is an extension of the standard Monte Carlo integration (MCI). It has been proved that SMCI is statistically more accurate than the standard MCI. Fine-tuning methods based on first-order and semi-second-order SMCI methods are formulated. The numerical experiments demonstrate that the proposed fine-tuning methods are superior to the conventional method in terms of both training and generalization errors.