Have literally spent the last couple of days trying to understand few shot learning for a university project and haven't been understanding it at all until this video. Great explanation, thank you so much!
does this mean if you train the model using a particular support set, then in testing, if you add a new classes to the support set (never seen before), the model would still be able to find that a query belongs to the new class even though it was never trained using that new class in the support set?
Why don't the similarities add up to 1? Aren't the classes mutually exclusive? Or is it not about the classes being mutually exclusive, but the fact the sample and the input overlap like a drawing and its reference. So that the an input image can be like other images even if that's not the correct class.
It depends on how they are computed. If they are the outputs of Softmax, then they add up to 1. If they are computed by the Siamese network, then they don't.