Efficient Feature Learning Using Perturb-and-MAP

Perturb-and-MAP [1] is a technique for efficiently drawing approximate samples from discrete probabilistic graphical models. These samples are useful for both characterizing the uncertainty in the model, as well as learning its parameters. In this work, we show that this same technique is effective at learning features from images using graphical models with complex dependencies between variables. In particular, we apply this technique in order to learn the parameters of a latentvariable model, the restricted Boltzmann machine, with additional higher-order potentials. We also use it in a bipartite matching model to learn features that are specifically tailored to tracking image patches in video sequences. Our final contribution is the proposal of a novel method for generating perturbations.