关键词: 受限玻尔兹曼机, 稀疏受限玻尔兹曼机, 竞争学习, 稀疏表示, 神经元

Abstract: Abstract: In order to resolve the difficulty of adaptive sparse of Sparse Restricted Boltzmann Machine, a new sparse mechanism is designed based on competition learning. Given the competition learning, the nearest cosine similarity measure was proposed for selection of RBM optimal neuron. The sparse competition process was designed which make optimal neuron constantly adjust the sparsity of other hidden units during RBM training process. Based on the deep model training process, the competition sparsity was applied to Deep Boltzmann Machine (DBM) construction. The handwritten numeral recognition experiments show that the sparse mechanism has better performance in hidden unit sparsity and features effectiveness compared to other sparse factors, and no more sparse parameters.

Key words: Keywords: Restricted Boltzmann Machine, Sparse RBM, Competition Learning, Sparse Representation, neuron