Comparison with proxy-NCA

[pbamotra]

Hi authors,

How does this method compare to this paper from Google research?
https://arxiv.org/pdf/1703.07464.pdf

Pankesh

[zhirongw]

Google's paper is sampling based approach for NCA. In each optimization step, they sample a triplet.
Ours is mostly similar to the original NCA method, and we don't do any sampling at all.
The use of the memory makes this computationally possible.