1/ Excuse me, can I interest you in eliciting your beliefs as flexible probability distributions? No worries, we only need pairwise comparisons or rankings, no personal details. Led by **Petrus Mikkola** and joint with **Arto Klami**, to be presented soon at @neuripsconf.bsky.social #NeurIPS2024 - ThreadSky

About ThreadSky

lacerbi.bsky.social • 86 days ago

1/ Excuse me, can I interest you in eliciting your beliefs as flexible probability distributions? No worries, we only need pairwise comparisons or rankings, no personal details.

Led by **Petrus Mikkola** and joint with **Arto Klami**, to be presented soon at @neuripsconf.bsky.social #NeurIPS2024

Comments

lacerbi.bsky.social•86 days ago

2/ In this paper, we learn joint pdfs from preference data or rankings, modelled as normalizing flows. A natural application is *prior elicitation* from experts.

This is not an easy task. Normalizing flows are *slippery* (technical term) and training can fail spectacularly, without constraints.

lacerbi.bsky.social•86 days ago

3/ Motivated by the exponential noise random
utility model (RUM), we introduce an *empirical
functional prior* that makes learning possible through function-space maximum-a-posteriori (MAP).

lacerbi.bsky.social•86 days ago

4/ The novel functional prior works! As a fun example, we extracted Claude 3 Haiku's joint prior over housing features in California, and compared to the California housing dataset, finding many matches. (We also tested other cases with known ground truth.)

lacerbi.bsky.social•86 days ago

5/ See the preprint for more details: https://arxiv.org/abs/2410.08710

And come and talk to us at our poster at #NeurIPS2024 @neuripsconf.bsky.social , both Petrus and I will be there!

Wed, Dec 11, 11:00 PST -- Poster Session 1 East

franzubaeren.bsky.social•86 days ago

Do you have a suggestion on how eliciting belief densities might revise the traditional reinforcement learning methods e.g. numerical or binary feedback?

Posting Rules

Be respectful to others
No spam or self-promotion
Stay on topic
Follow Bluesky's terms of service

Comments

Posting Rules

Reply