AAM-Softmax is also called ArcFace and is proposed in paper: ArcFace: Additive Angular Margin Loss for Deep Face Recognition. This loss function is equivalent to CosFace except that there is additive *angular* margin penalty *m* between \(x_i\) and \(W_{yi}\).

The additive angular margin penalty is equal to the geodesic distance margin penalty in the normalised hypersphere.

It is formulated as:

We also can use this loss function in speaker verification task. In that task, we usually set s = 30, m = 0.2