Convolutional Restricted Boltzmann Machines for Feature Learning

In this thesis, we present a method for learning problem-specific hierarchical features specialized for vision applications. Recently, a greedy layerwise learning mechanism has been proposed for tuning parameters of fully connected hierarchical networks. This approach views layers of a network as Restricted Boltzmann Machines (RBM), and trains them separately from the bottom layer upwards. We develop Convolutional RBM (CRBM), an extension of the RBM model in which connections are local and weights are shared to respect the spatial structure of images. We switch between the CRBM and down-sampling layers and stack them on top of each other to build a multilayer hierarchy of alternating filtering and pooling. This framework learns generic features such as oriented edges at the bottom levels and features specific to an object class such as object parts in the top layers. Afterward, we feed the extracted features into a discriminative classifier for recognition. It is experimentally demonstrated that the features automatically learned by our algorithm are effective for object detection, by using them to obtain performance comparable to the state-of-the-art on handwritten digit classification and pedestrian detection.