Soumith Chintala: evaluation

Transcript context

…Already is a rage. Which I think is fairly misplaced in how people perceive it. People think synthetic data is some kind of magic wand that you wave and it's going to be amazing. Synthetic data is useful in neural networks right now because we as humans have figured out a bunch of symbolic models of the world or made up certain symbolic models because of human innate biases. So we've figured out how to ground particle physics in a 30 parameter model. And it's just very hard to compute as in it takes a lot of flops to compute, but it only has 30 parameters or so. I mean, I'm not a physics expert, but it's a very low rank model. We built mathematics as a field that basically is very low rank. Language, a deep understanding of language, like the whole syntactic parse trees and just understanding how language can be broken down and into a formal symbolism is something that we figured out. So we basically as humans have accumulated all this knowledge on these subjects, either synthetic, we created those subjects in our heads, or we grounded some real world phenomenon into a set of symbols. But we haven't figured out how to teach neural networks symbolic world models directly. The only way we have to teach them is generating a bunch of inputs and outputs and gradient dissenting over them. So in areas where we have the symbolic models and we need to teach all the knowledge we have that is better encoded in the symbolic models, what we're doing is we're generating a bunch of synthetic data, a bunch of input output pairs, and then giving that to the neural network and asking it to learn the same thing that we already have a better low rank model of in gradient descent in a much more over-parameterized way. Outside of this, like where we don't have good symbolic models, like synthetic data obviously doesn't make any sense. So synthetic data is not a magic wand where it'll work in all cases in every case or whatever. It's just where we as humans already have good symbolic models off. We need to impart that knowledge to neural networks and we figured out the synthetic data is a vehicle to impart this knowledge to. So, but people, because maybe they don't know enough about synthetic data as a notion, but they hear, you know, the next wave of data revolution is synthetic data. They think it's some kind of magic where we just create a bunch of random data somehow. They don't think about how, and then they think that's just a revolution. And I think that's maybe a gap in understanding most people have in this hype cycle. Yeah, well, it's a relatively new concept, so. Oh, there's two more that I'll put in front of you and then you can see what you respond. One is, you know, I have this joke that it's, you know, it's only synthetic data if it's from the Mistral region of France, otherwise it's just a sparkling distillation, which is what news research is doing. Like they're distilling GPT-4 by creating synthetic data from GPT-4, creating mock textbooks inspired by Phi 2 and then fine tuning open source models like Llama. And so I don't know, I mean, I think that's, should we call that synthetic data? Should we call it something else? I don't know.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.