James Zou
speaker
371 appearances
1 recordings
1 series
first heard Dec 2024
last heard Dec 2024
James Zou’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
And what we wanted to do is to see can we actually use these similar ideas to apply that to like approaching language model.
So I I already mentioned ESM in the context of the virtual lab.
It's actually one of the tools that's used by the virtual lab agents, but it's basically like a protein language model that's trained over this universe of different protein sequences.
The idea there is that you can view actually each amino acid sequence of a protein is speaking like a sentence where the characters or the the words are now different in the residues of the amino acid.
And then people would train similar kind of transformers using masking.
right on top of those sentences of proteins.
And that's shown to be a very powerful approach because uh similar to Hubb's human language, right, there's a lot of
hidden grammars and things with the proteins and these protein language models like ESM, because they've been trained over evolutionary scale data, they can sort of capture those intrinsic and hidden grammars of proteins in its embeddings.
And this is where also then starts to diverge a bit from the human models, because the human model, we sort of have a pretty good idea of how to interpret what are the concepts that are learned by a a GPT style human language model, right?
Maybe we know the concept like Golden Gate Bridge, we can see other particular neurons in a transformer that corresponds to the Golden Gate Bridge.
But where it's sort of fascinating with the protein language model is that there there are some concepts that we know.
Maybe I know a particular kind of protein domain.
I can see, okay, does that show up somewhere in that protein language model?
But I would say actually most of the concepts are probably not known to us, right?
They're sort of uh we don't have a good description about what are these concepts.
And this is where I see these interpretation of a protein language model not just being like an exercise in mechanistic interpretability, but actually as a sort of an exercise in scientific discovery, because I think there's actually a gold mine of
new concepts that potentially learned by this protein language model training over these evolutionary scale data that we just do not actually have good human analogs for.
And if we can actually go in and look at different layers and use SAEs to tease apart and here are the different new concepts and then validate those concepts, then that can actually become like a really a gold mine of new biological insights that the model can teach us.
Yeah, I think especially in these more specialized domains like proteins or biology
No there's a very clear limitation of what we have
Showing 301–320 of 371 · page 16 of 19
← Previous
Next →