I am trying to understand a paper that uses a siamese network for authentication.

In this paper (see Figure 4), they have two parts: mobile device and cloud. Negative samples are stored in the cloud and positives are on a mobile phone. They state that while training, they extract the feature vector of negative examples in the cloud and the feature vector of positive examples in the mobile phone.

I do not understand how this can be possible, because as far as I understood, positive and negative examples are used to find shared model weight. For me, it means that we have one set of weights (= one weight matrix) for each layer (not one for positive and one for negative samples). Therefore, how we train in that way. I think we need to give 2 inputs (a pair) at the same time to the network, but how these two parts can have shared weights. Maybe we can't call it siamese network now, I don't know.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.