Comment by cs702

Yes, it needs prototypes (i.e., possible neighbors) -- but it doesn't need any kind of pretraining to learn to map samples to a feature space.

It uses gzip instead of a pretrained model to map samples to the feature space.