An Efficient Normalized Restricted Boltzmann Machine for Solving Multiclass Classification Problems

Multiclass classification based on unlabeled images using computer vision and image processing is currently an important issue. In this research, we focused on the phenom-ena of constructing high-level features detector for class-driven unlabeled data. We proposed a normalized restricted Boltzmann machine (NRBM) to form a robust network model. The proposed NRBM is developed to achieve the goal of dimensionality reduc-tion and provide better feature extraction with enhancement in learning more appropriate features of the data. For increment in learning convergence rate and reduction in complexity of the NRBM, we add Polyak Averaging method when training update parameters. We train the proposed NRBM network model on five variants of Modified National Institute of Standards and Technology database (MNIST) benchmark dataset. The conducted experiments showed that the proposed NRBM is more robust to noisy data as compared to state-of-art approaches.