Adaptive structure learning method of deep belief network using neuron generation–annihilation and layer generation

Recently, deep learning is receiving renewed attention in the field of artificial intelligence. Deep belief network (DBN) has a deep network architecture that can represent multiple features of input patterns hierarchically, using pre-trained restricted Boltzmann machines (RBMs). Such deep network architectures enable extremely high classification accuracy in many tasks compared to previous methods. However, determining various parameters to design effective deep network architectures is a difficult task even for experienced designers, since traditional RBM and DBN cannot change their network structure during the training. The adaptive structure learning method has been previously proposed for finding the optimum number of hidden neurons in multilayered neural networks. The method employs the neuron generation–annihilation algorithm by observing the variance of weight decays. We develop the adaptive structure learning method of RBM and DBN using the neuron generation–annihilation and layer generation algorithm by observing the variance of some parameters. The effectiveness of our proposed model was verified by tenfold cross-validation on benchmark data sets CIFAR-10 and CIFAR-100. The adaptive DBN achieved the highest classification accuracy (97.4% for CIFAR-10, 81.2% for CIFAR-100) among several latest DBN- and CNN-based methods.