Accelerating AI Performance using Anderson Extrapolation on GPUs