Angular Prototypical Loss

This text is study for arxiv.org/pdf/2003.11982.pdf

 

Prototypical loss

• Each mini-batch contains a support set S and query set Q
• Query is M-th utterance from every speaker
• The prototype (centroide) is [1]
• Squared euclidian distance is used as the distance metric [2]
• Each query example is classified against N speakers based on softmax [3]

[1]

 

[2]

[3]

 

 

 

Generalized end-to-end (GE2E) loss

• Every utterance in the batch except the query itself is used to form centroids [1]
• The similarity matrix is defined as scaled cosine similarity between embeddings and all centroids [2]
• The final GE2E loss is [3]

 

 

[1]

[2]

[3]

 

Angular Prototypical loss

• Use the same batch formation as the original prototpical loss, reserving one utterance from every class as query
• This can make possible to exactly mimic the test scenario during training
• Cosine-based similarity metric [1]
• The objective function is similar with prototypical loss

[1]