Graph-based transformation manifolds for invariant pattern recognition with kernel methods

We present here an approach for applying the technique of modeling data transformation manifolds for invariant learning with kernel methods. The approach is based on building a kernel function on the graph modeling the invariant manifold. It provides a way for taking into account nearly arbitrary transformations of the input samples. The approach is verified experimentally on the task of optical character recognition, providing state-of-the-art performance on harder problem settings.