GPU parallel implementation of the approximate K-SVD algorithm using OpenCL

Training dictionaries for sparse representations is a time consuming task, due to the large size of the data involved and to the complexity of the training algorithms. We investigate a parallel version of the approximate K-SVD algorithm, where multiple atoms are updated simultaneously, and implement it using OpenCL, for execution on graphics processing units (GPU). This not only allows reducing the execution time with respect to the standard sequential version, but also gives dictionaries with which the training data are better approximated. We present numerical evidence supporting this somewhat surprising conclusion and discuss in detail several implementation choices and difficulties.