Where are the optimization potential of machine learning kernels?

In this paper we investigate the research directions for the optimization of machine learning (ML) kernels on heterogeneous architectures. Heterogeneous computing systems are systems that use more than one kind of processor or cores. Machine learning is the sub field of computer science that "gives computers the ability to learn without being explicitly programmed". Efficient ML kernels on heterogeneous systems are a challenging task and relatively new in the research community. As a concrete realization in order to "measure" the difficulty of the task, we experiment with two well known ML kernels, namely Scikit-learn and MLlib and we realize some experiments on a FPGA board with the Xilinx technology. The conclusions we draw from our study are that FPGA implementations of mathematical procedures (BLAS) are possible to accelerate ML kernels but with limited performance gains. We also suggest to put an extreme effort on coarse grained approaches (at the procedure level) for the optimization of ML kernels because we guess that the potential for optimization is located at this level.