NeurIPS 2020

GPU-Accelerated Primal Learning for Extremely Fast Large-Scale Classification


Meta Review

The paper exhaustively discusses improvements of the TRON algorithm (Trust Region Newton Algorithm), an efficient solver for L2 primal problems, to benefit from both CPUs and GPUs. The improvements are based on increasing parallel processing on CPUs and GPUs, decoupling sequential dependencies of variables, and minimizing the frequency of large memory transfers between CPU and GPU. Overall, this is a solid paper. The reviewers were pointing out that some claims were not justified or misleading, however, the authors succeeded to deliver either an appropriate justification or promised to revise a given claim.