Improved Gaussian–Bernoulli restricted Boltzmann machine for learning discriminative representations

Highlights：

• We focus on exploring the valuable information from within the data to guide the unsupervised training process of the GBRM, rather than relying on external labels.

• The informative interpoint affinity information can be simultaneously introduced into the data process and reconstruction process of CD learning, which can take full advantage of the guiding function of those information, and facilitate in learning discriminative representations.

• The proposed abGRBM model can not only utilize the GRBM’s powerful latent representation learning capabilities for real-valued data, but also simultaneously transform the original data into anther space with the property of better separability.

• Experiments conducted on several Microsoft Research Asia Multimedia (MSRA-MM) image datasets show that the representations learned by our abGRBM model are more conducive to improving clustering and classification performance than those of the traditional GRBM.

摘要

•We focus on exploring the valuable information from within the data to guide the unsupervised training process of the GBRM, rather than relying on external labels.•The informative interpoint affinity information can be simultaneously introduced into the data process and reconstruction process of CD learning, which can take full advantage of the guiding function of those information, and facilitate in learning discriminative representations.•The proposed abGRBM model can not only utilize the GRBM’s powerful latent representation learning capabilities for real-valued data, but also simultaneously transform the original data into anther space with the property of better separability.•Experiments conducted on several Microsoft Research Asia Multimedia (MSRA-MM) image datasets show that the representations learned by our abGRBM model are more conducive to improving clustering and classification performance than those of the traditional GRBM.