Daniel Jeffries (Unknown)

I had him on the show.

Can you explain what this paper was about?

It's a fascinating thing about neural networks that, as you say, you can actually randomly assign labels, not even in a consistent way.

And the neural network will still get 100% accuracy on the train set.

And I guess that's because if you think about it, the neural networks have enough flexibility to kind of place decision boundaries around each of the individual training examples so they can basically memorize your training set a bit like a zip file or something like that.

So there's the spectrum, isn't there, that you can, on the one side, you can memorize examples.

And then going up a little bit on the spectrum, you can memorize features.

So like these weird blue dots that you're talking about.

And then somewhere up on the spectrum, you end up with robust, out-of-domain generalization features that actually represent the thing that you want.

So I guess the question is, how do you know the difference between it just memorizing examples versus non-robust features?

Yeah.

So how did you go about, you know, mitigating the robustness problem in the paper?

Yeah.

And I'm thinking about the different approaches.

I mean, hypothetically, you could preprocess the data before it even goes into the model.

You could change the optimization algorithm of the model itself, or given a trained model, you could kind of like, you know, robustify it.

So you're talking about the middle option where you're actually changing the optimization algorithm.

Yes.

Yeah, I mean, I'm interested in the different types of features, as you were just alluding to.

So Randall Belistriero yesterday, he had this paper which was talking about the difference between reconstructive methods like, you know, let's say a masked autoencoder versus, you know, a self-supervised contrastive image representation learning model.