Learning from Few Samples: Transformation-Invariant SVMs with Composition and Locality at Multiple Scales