Elan Barenholtz
speaker
2,007 appearances
4 recordings
1 series
first heard Feb 2025
last heard Jun 2025
Elan Barenholtz’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsNo recordings in the last 12 months.Older appearances are listed below; set an alert to hear about the next one.
Appearances
Um but you send it through this latent space uh and then recover it uh in its original form.
What what you could do is once you've got that latent space, you can then translate
from one embedding to a completely different embedding.
Mm-hmm.
This is a new paper.
This is a new paper, yes.
Right, right, right.
This sh this rocked my world.
Uh because what they were what they're arguing is that there in some ways is this uh underlying uh universal
s uh structure of language that's captured in this latent space.
And so even though if you have a radically different embedding in one line, you know, they didn't do it across different languages.
That's one of the uh projects I'm doing right now uh is to see if you can do this across, say, English and Spanish.
Um, even for a language that's trained exclusively on English and then another like another models trained exclusively on Spanish, can you guess the Spanish just from uh finding these kind of universal uh pr uh structure?
across uh these two different models.
So the way to think about it is that what you're doing is creating uh an another embedding, another another this this latent space
w where you're gonna be able to send in uh a message in English and then uh based on
the station and then again do the same thing for Spanish.
And then what you're not you're never gonna show any mon no model is going to ever see a pair.
of English and Spanish.
Instead, what you're gonna learn is that there's some way to get from uh lan you're gonna end up being able to get from English to Spanish without ever seeing the actual translation.
Showing 241–260 of 2,007 · page 13 of 101
← Previous
Next →