Learning framework of multimodal Gaussian–Bernoulli RBM handling real-value input data

The conventional Gaussian–Bernoulli restricted Boltzmann machine (GBRBM), which is a RBM model for processing real-valued data, presumes single Gaussian distribution for learning real numbers. However, a single distribution is not able to effectively reflect complex data in many cases of real applications. In order to overcome this limitation, Gaussian mixture model (GMM) based RBM is proposed. As a learning mechanism for the proposed model, an energy function handling multi-modal distribution is provided. Then, a memetic algorithm (MA) was applied in order to train the proposed framework more accurately in real-valued input data. In order to show the effectiveness of the proposed framework, the method is applied to image reconstructions. The experiments show that the proposed framework provides more valid results than the other RBM based models in reconstruction error. Through the experiment results, it is concluded that the proposed framework is able to apply real-valued input data extensively and reduce difficulties of learning parameters by capturing the characteristics of real-value input data using GMM.