Andrew Ilyas

I won't go into detail about that here.

But the interesting thing is that we found that TRAQ was a really effective tool for finding the most important training examples for a given test prediction.

Yeah, that's a fantastic question.

And actually, one of the inspirations for developing the whole data model framework is that we wanted to understand there's the spectrum between very exemplar-based prediction of, I say this is a dog because I've seen a similar dog before, and the abstract high-level feature learning.

And we wanted to actually go about testing that hypothesis.

We wanted to see, at the test set level, how many features are learned locally or on an exemplar basis versus how many are these abstract features.

And what we found is that there is a good core of what you might call abstract features.

It's hard to actually pin them down as abstract features.

But at the very least, there's a good core of images or of examples that are learned very robustly.

And so if you look at what the data attributions for these examples look like, they're very dense in data space.

They're abstracting from a lot of different training examples.

but there is sort of a very significant tale of examples that are learned that seem to be learned at least on a very exemplar, like in a very exemplar like way.

So there are a bunch of examples for which you can remove like, you know, 10, 20 out of 50,000 training samples and flip a model's prediction on those training examples, on those test examples, sorry.

Yeah, so I haven't had a great, I don't have a great conception of what that is.

I think it very much also depends on the context in which you're talking.

I think when people think about, like you were saying, there's this notion of robust versus non-robust features in the causal sense.

There's robust versus non-robust features in the adversarial sense.

Now, the thing I was just talking about is a notion of robust versus non-robust features in the training data sense.

And so I think for each of those senses, you can sort of define what it means for a feature to be robust.

But I don't think we really have a great conception of what exactly a feature is.