Scores of emerging and domain-specific applications need the ability to acquire and augment new knowledge from offline training-sets and online user interactions. This requires an underlying computing platform that can host machine learning (ML) kernels. This in turn entails one to have efficient implementations of the frequently-used ML kernels on state-of-the-art multicores and many-cores, to act as high-performance accelerators. Motivated by this observation, this paper focuses on one such ML kernel, namely, K Nearest Neighbor (KNN), and conducts a comprehensive comparison of its behavior on two alternate accelerator-based systems: NVIDIA GPU and Intel Xeon Phi (both KNC and KNL architectures). More explicitly, we discuss and experimentally evaluate various optimizations that can be applied to both GPU and Xeon Phi, as well as optimizations that are specific to either GPU or Xeon Phi. Furthermore, we implement different versions of KNN on these candidate accelerators and collect experimental data using various inputs. Our experimental evaluations suggest that, by using both general purpose and accelerator specific optimizations, one can achieve average speedups ranging 0.49x-3.48x (training) and 1.43x-9.41x (classification) on Xeon Phi series, compared to 0.05x-0.60x (training), 1.61x-6.32x (classification) achieved by the GPU version, both over the standard host-only system.