Daniel Jeffries (Unknown)

So how does that actually work in a production setting?

Very cool.

So we should go to your data modeling paper first.

So I think you wrote this about two years ago and you were studying essentially how algorithms combine together with data to yield model predictions.

Can you kind of explain that paper?

Yeah.

So there are some similar things that have been done before.

I mean, there was, I think, influence functions and Shapley values and so on.

How does it kind of relate to those?

Very interesting.

I wonder, how is it related to machine teaching?

I mean, I found out about machine teaching a few years ago.

And as I understand it, it's a way to kind of find the minimal data set.

So some kind of meta process where we carve down the data set to find the minimum size one that will still do well on the problem.

Very cool.

So as I understand it, data modeling is like a black box method where we have a surrogate model.

And the model is a function of the algorithm, an input example, and some training subset.

And it's predicting the output of the training and evaluation of the model.

So the first thought is these models are really complex.

And we're just building a linear surrogate function.