Transformer Memory as a Differentiable Search Index: memorizing thousands of random doc ids works!?
Neural Search Talks — Zeta Alpha
00:00
Using Hits at One and Hits at Ten
There are a number of documents in the full collections around 230K, right? So it's what, 40 times or so to go to the full MS Marko size. And you can anticipate there has to be capacity issues at some point somewhere. I think just for the sake of demonstrating that this can work, it makes sense to start with something reasonably sized. This is way, way larger than they're looking at here.
Play episode from 26:09
Transcript
