Masked Multi Self Attention In Deep Learning