Learning a dictionary of deformable patches using GPUs

We propose a simple method for learning a dictionary of deformable patches for simultaneous shape recognition and reconstruction. Our approach relies on two key innovations - introducing a pre-defined set of transformations on patches to enrich the search space, and designing a parallel framework on Graphical Processors (GPUs) for matching a large number of deformable templates to a large set of images efficiently. We illustrate our method on two handwritten digit databases - MNIST and USPS, and report state-of-art recognition performance without using any domain-specific knowledge on digits. We briefly show that our dictionary has many desirable properties: it includes intuitive low- and mid-level structures, it is sufficient to synthesize digits, it gives sparse representations of digits, and contains elements which are useful for discrimination. In addition, we are the first dictionary learning method to report good results when transferring the learned dictionary between different datasets.