137 - Nearest Neighbor Language Modeling and Machine Translation, with Urvashi Khandelwal
NLP Highlights
00:00
Canon MT Paper - What Are the Most Exciting Open Questions?
I think we highlighted the interesting aspects. We were able to sort of reinforce the domain adaptation finding that a single model can be useful in multiple domains by simply adding a domain-specific data store. I would encourage people to think of this general framework in settings where explicitly drawing associations might be helpful, probably in sequential decision-making problems. But moving forward, thinking about it less as a token level retrieval mechanism, but more as a high level generalization through memorization, like drawing associations as a framework could open up a wealth of possibilities for how we can use this in the future.
Transcript
Play full episode
