Improving the Neural GPU Architecture for Algorithm Learning