An overview of SCG and how gradient accumulation can help optimization when using minibatches for training neural nets.