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
Because what the model is gonna learn is what's common across these two representations.
What's true for both the Spanish abetting and the English abetting, that there's some sort of underlying latent structure
structure that's true of both and that that captures something more universal about language.
Now again, they didn't do it for different languages.
They just did it for different uh embeddings of English, um, but very different embeddings because they were trained on completely different models.
If you looked at them, if you just looked at this sort of vector representation, you took a vector uh representation of the word dog in one and a vectorization of the word dog in the other, they're completely numerically not there's no similarity.
You can never spot the similarities if you just looked even them pairwise.
But if they do this kind of reconstruction and then ask the model,
to be able to reconstruct not the the the in the original uh the the original embedding space, but go and reconstruct in the
uh the other embedding space, it's able to actually do this.
And so by doing that, it's by training it to do that without ever seeing any pairs, it's able to sort of learn the translation between one representation and another representation.
Um
what the the what what this opened up to me is is the possibility that we could think about the exact same kind of latent space in the brain, uh and possibly in uh in in artificial intelligence models, between say the perceptual world
and the linguistic world, that there is uh some embedding of uh uh how the physical world is structured.
Uh we understand like an uh think about an animal, a non linguistic animal certainly has idea of objects and objects in relation to other objects, uh objects in proximity to other objects, moving around those objects.
My dog, uh who was just barking in the background, knows what doors are and she can go scratch at it and she knows it opens up.
She certainly isn't able to express that linguistically, but she has this concept and she's able to think about that.
She's able in some ways to reason about that.
My suspicion is that that probably is done maybe even auto-aggressively, but we'll leave that aside for now.
The main point is that there is some representation in
Showing 261–280 of 2,007 · page 14 of 101
← Previous
Next →