Efficient Self Attention In Deep Learning