Convolutional restricted boltzmann machines for feature learning

In this thesis, we present a probabilistic generative approach for learning hierarchical structures of spatially local features, effective for visual recognition. Recently, a greedy layerwise learning mechanism has been proposed for training fully-connected neural networks. This mechanism views each of the network's layers as a Restricted Boltzmann Machines (RBM), and trains them separately and bottom-up. We develop Convolutional RBM (CRBM), in which connections are local and weights are shared to respect the spatial structure of images. We train a hierarchy of visual feature detectors in layerwise manner by switching between the CRBM models and down-sampling layers. Our model learns generic gradient features at the bottom layers and class-specific features in the top levels. It is experimentally demonstrated that the features automatically learned by our algorithm are effective for visual recognition tasks, by using them to obtain performance comparable to the state-of-the-art on handwritten digit classification and pedestrian detection.