(a) I think “causal representation learning” is too vague, this overview (https://arxiv.org/pdf/2102.11107.pdf) talks about a lot of different problems I would consider fairly unrelated under this same heading.
Yes, you’re right. I had found this, and other reviews by similar authors. In this one, I was mostly thinking of section VI (Learning causal variables) and its applications to RL (section VII-E). Perhaps section V on causal discovery is also relevant.
(b) I would try to read “classical causal inference” stuff. There is a lot of reinventing of the wheel (often, badly) happening in the causal ML space.
Probably there is, I have to get to speed on quite a few things if I get the grant. But thanks for the nudge!
(c) What makes a thing “causal” is a distinction between a “larger” distribution we are interested in, and a “smaller” distribution we have data on. Lots of problems might look “causal” but really aren’t (in an interesting way) if formalized properly.
I think this makes sense. But part of the issue here is that AI will probably change things that we have not foreseen, so it could be good to take this point of view, in my opinion. Do you have `interesting’ examples of not-causal problems?
Classical RL isn’t causal, because there’s no confounding (although I think it is very useful to think about classical RL causally, for doing inference more efficiently).
Various extensions of classical RL are causal, of course.
A lot of interesting algorithmic fairness isn’t really causal. Classical prediction problems aren’t causal.
However, I think domain adaptation, covariate shift, semi-supervised learning are all causal problems.
---
I think predicting things you have no data on (“what if the AI does something we didn’t foresee”) is sort of an impossible problem via tools in “data science.” You have no data!
I think value learning might be causal because human preferences cannot be observed, and therefore can act as a confounder, similar to the work in
Zhang, J., Kumor, D., Bareinboim, E. Causal Imitation Learning with Unobserved Confounders. In Advances in Neural Information Processing Systems 2020.
At least that was one of my motivations.
I think predicting things you have no data on (“what if the AI does something we didn’t foresee”) is sort of an impossible problem via tools in “data science.” You have no data!
Sure, I agree. I think I was quite inaccurate. I am referring to transportability analysis, to be more specific. This approach should help in new situations where we have not directly trained our system, and in which our preferences could change.
Hi Ilya! Thanks a lot for commenting :)
Yes, you’re right. I had found this, and other reviews by similar authors. In this one, I was mostly thinking of section VI (Learning causal variables) and its applications to RL (section VII-E). Perhaps section V on causal discovery is also relevant.
Probably there is, I have to get to speed on quite a few things if I get the grant. But thanks for the nudge!
I think this makes sense. But part of the issue here is that AI will probably change things that we have not foreseen, so it could be good to take this point of view, in my opinion. Do you have `interesting’ examples of not-causal problems?
Thanks again!
Classical RL isn’t causal, because there’s no confounding (although I think it is very useful to think about classical RL causally, for doing inference more efficiently).
Various extensions of classical RL are causal, of course.
A lot of interesting algorithmic fairness isn’t really causal. Classical prediction problems aren’t causal.
However, I think domain adaptation, covariate shift, semi-supervised learning are all causal problems.
---
I think predicting things you have no data on (“what if the AI does something we didn’t foresee”) is sort of an impossible problem via tools in “data science.” You have no data!
I think value learning might be causal because human preferences cannot be observed, and therefore can act as a confounder, similar to the work in
Zhang, J., Kumor, D., Bareinboim, E. Causal Imitation Learning with Unobserved Confounders. In Advances in Neural Information Processing Systems 2020.
At least that was one of my motivations.
Sure, I agree. I think I was quite inaccurate. I am referring to transportability analysis, to be more specific. This approach should help in new situations where we have not directly trained our system, and in which our preferences could change.