Modeling pixel means and covariances using factorized third-order boltzmann machines

摘要: Learning a generative model of natural images is useful way extracting features that capture interesting regularities. Previous work on learning such models has focused methods in which the latent are used to determine mean and variance each pixel independently, or hidden units covariance matrix zero-mean Gaussian distribution. In this work, we propose probabilistic combines these two approaches into single framework. We represent image using one set binary image-specific separate mean. show approach provides framework for widely simple-cell complex-cell architecture, it produces very realistic samples extracts yield state-of-the-art recognition accuracy challenging CIFAR 10 dataset.