Infinite Fuzzy Restricted Boltzmann Machines for the Enhancement of Fuzzy Neural Networks

Nowadays, fuzzy restricted Boltzmann machines (FRBMs) enjoy wide attention in deep learning and fuzzy systems. As a fuzzy neural network, FRBM is the extension of the classic RBM. However, FRBM has inherent defects because it relies heavily on the size of hidden units for feature extraction. In this paper, a novel infinite FRBM (IFRBM) is proposed by extending the hidden units from fixed parameters to infinite ones. The structure of the FRBM is further improved guaranteed by a novel infinite energy function without specifying the hidden elements. Specifically, the proposed model can increase generated hidden units adaptively during training when encountering different training data. Comparative experiments demonstrate that our IFRBM can obtain competitive performance than the state-of-the-art model in terms of feature extraction and image classification. Besides, the proposed model tends to more concise network structure while not requiring the tuning of hidden hyper-parameter or additional expert experience.