Multiplicative Attention

• 
• $f_{att}(h_i,s_j)=h_i^T{W}_a{s}_j$
• Since Additive Attention performs better for scale, use a factor Scaled Dot Product Attention