"Conditional Causal Discovery"

(don't be fooled by the title :D )

openreview.net/forum?id=6IY...

"Estimating the Probabilities of Rare Outputs in Language Models"

arxiv.org/abs/2410.13211

"Stochastic Optimization with Optimal Importance Sampling"

arxiv.org/abs/2504.03560

Posting a few nice importance sampling-related finds

"Value-aware Importance Weighting for Off-policy Reinforcement Learning"

proceedings.mlr.press/v232/de-asis...

Wish I had found this when I got started : P (or, maybe not)

You are free to estimate your integrals by approximating p without weighting.
I am free to clean my room by throwing everything inside the closet.

Do I have a reading problem

Thank you, I will check the thesis !
Regarding the paper: I was particularly interested in what they mention as future work as far as I understand

nice idea.. seen it before 😉

Other toy example visualizing the effect of tail dependence in a ratio

Sometimes a ratio estimate will be better than another one even if the individual num/den estimators perform similarly-because a correlation between num/denom induces some error cancellation.
So looking only at k-hat values doesn't fully explain the diff. in performance of two competing estimators.

When the target quantity of interest is a ratio (like posterior predictives), we actually have 2 sets of weights - for num. and denom.
The reliability of the two estimators, i.e., of num. and denom., can be *individually* assessed with the k-hat diagnostic of Vehtari et al (2024, JMLR).

More details: consider the general problem of estimating expectations. Most of these (e.g. in Bayesian inf.) involve an intractable normalizing constant so ratios of estimators are used in practice.
Intuitively, reliability of an IS estimator depends on the behavior of the density ratios (weights)

Interested in estimating posterior predictives in Bayesian inference? Really want to know if your approximate inference "is working"?
Come to our poster at the NeurIPS BDU workshop on Saturday - see TL;DR below.

These things I usually don't see much discussed. I guess it's not that interesting to look into anything else than the individual variances per dimension (or their average / sum).

A bonus of the nice notes by François Portier:
explicit effect of dimension on the error !
@nolovedeeplearning.bsky.social

The above early reference of 1966 starts explicitly calling it "adaptive" and mentions the idea appeared already in 1954, but did not receive much attention.

- artowen.su.domains/mc/Ch-var-ad...
- the notes on Monte Carlo here: sites.google.com/site/fportie...
- arxiv.org/abs/2102.05407
All have different (complementary) material. The basic concept is super general (maybe too general I suspect @nolovedeeplearning.bsky.social may argue :D ).

www.tandfonline.com/doi/full/10....