Sequential Attention Making models leaner and fast without sacrificing accuracy (research.google)