Deep Generative Deconvolutional Image Model.pdfVIP

A Deep Generative Deconvolutional Image Model Yunchen Pu Xin Yuan Andrew Stevens Chunyuan Li Lawrence Carin Duke University Bell Labs Duke University Duke University Duke University Abstract A deep generative model is developed for rep- resentation and analysis of images, based on a hierarchical convolutional dictionary-learning framework. Stochastic unpooling is employed to link consecutive layers in the model, yielding top-down image generation. A Bayesian support vector machine is linked to the top-layer features, yielding max-margin discrimination. Deep de- convolutional inference is employed when test- ing, to infer the latent features, and the top-layer features are connected with the max-margin clas- sifier for discrimination tasks. The model is effi- ciently trained using a Monte Carlo expectation- maximization (MCEM) algorithm, with imple- mentation on graphical processor units (GPUs) for efficient large-scale learning, and fast testing. Excellent results are obtained on several bench- mark datasets, including ImageNet, demonstrat- ing that the proposed model achieves results that are highly competitive with similarly sized con- volutional neural networks. 1 Introduction Convolutional neural networks (CNN) (LeCun et al., 1989) are effective tools for image and video analysis (Chatfield et al., 2014; Krizhevsky et al., 2012; Mnih et al., 2013; Ser- manet et al., 2013). The CNN is characterized by feedfor- ward (bottom-up) sequential application of convolutional filterbanks, pointwise nonlinear functions (e.g., sigmoid or hyperbolic tangent), and pooling. Supervision in CNN is typically implemented via a fully-connected layer at the top of the deep architecture, usually with a softmax clas- sifier (Ciresan et al., 2011; He et al., 2014; Jarrett et al., 2009; Krizhevsky et al., 2012). A parallel line of research concerns dictionary learning Appearing in Proceedings of the 19th International Conference on Artificial Intelligence and Statistics (AISTATS) 2016, C