Classification layer

[golunovas]

Thanks for the great idea of using classification layer for cosine distance computation! Could you please elaborate on how you train that layer? As I understand you initialize weights with average feature vectors for each class and what do you do with biases? Do you use them for training?