Research Of Deep Learning Method Based On Restricted Boltzmann Machines

Deep learning is an important research direction in machine learning field and has developed well in recent years. The main idea of deep learning is building hierarchical structures of deep network and applying effective training algorithms to extract higher level features and accomplish unsupervised or supervised tasks. Deep learning methods can be regarded as a further development of artificial neural network which solved the training problem of multilayer neural network and they also have been inspired by neural science findings and other machine learning theories and technologies. In the mainstream of deep learning models, Restricted Boltzmann Machines is a special and significant component composed of two layers'units. By means of its statistical properties, the deep structure can be trained by a layer-wise procedure first, and then fine-tuned by a training algorithm of traditional multilayer neural network. This kind of combined training simplifies the learning and improves the efficiency of feature extraction on unlabeled data. The main research contents are as follows:1. There is a disadvantage of Restricted Boltzmann Machines that the units in the same layer have no connections with each other. This may cause the loss of inner information of the input data. Taking this into consideration, we proposed a new RBM and DBN structure. Glia cells are special neural cells exist in human brains, and connected with common neurons. They can adjust the states of neurons and send signal to other glia cells. We add the glia chains into the RBM by defining new activation rules, and build new deep structures trained by an improved algorithm, to increase the learning speed of deep network and enhance the performance of data feature extraction.2. With the added glia chains, the units of RBM in the same layer can transmit signals but the activation of the units are not regularized so that the abstracted features are not distinguishing. So we introduce the theory of self-organizing maps in neural network and construct new model of RBM with a new training algorithm to further improve the learning efficiency and obtain better image features.3. Deep learning methods are widely used in different applications. In this research we mainly focus on the improvement of RBM's model structure and its application in image classification. By performing a lot of experiments on three image datasets, we tested and verified the classification accuracy and converge speed of the new model, managed to adjust the parameters to optimal and made it adapted to image classification and recognition tasks, to reach its optimum performance.