Talk:Evidence lower bound
Add topic | This article is rated C-class on Wikipedia's content assessment scale. It is of interest to the following WikiProjects: | ||||||||||||||||||||
| |||||||||||||||||||||
This article needs images
[edit]maybe inspired by https://www.youtube.com/watch?v=HxQ94L8n0vU Biggerj1 (talk) 19:27, 17 October 2021 (UTC)
First appearance of ELBO
[edit]I wondered which publication introduced this idea. The first appearance I can think of is in the Neal & Hinton 1998 paper (which is actually from 1993, see the Hinton & Zemel paper). Maybe someone knows of an earlier mention of this concept? Apoptheosis (talk) 09:33, 5 November 2023 (UTC)
- i don't know where the concept first appeared. i do, however, know where the particular name "ELBO" first appeared, because i made it up. Jonmcauliffe (talk) 07:00, 18 December 2023 (UTC)
- Yes, Neal & Hinton 1998 is the standard reference for variational EM. However, the use of an amortized inference network q was introduced by Kingma & Welling 2014, which is the standard reference for VAEs. Eclecticos (talk) 08:32, 16 November 2025 (UTC)
People are adding their own research here
[edit]This is inappropriate for an encyclopedia. Stuff about delta method, weighting is to important enough to qualify for this entry.
Wikipedia:No original research EitanPorat (talk) 01:52, 13 March 2024 (UTC)
Learning
[edit]There should be some more info here about how p and q are trained in practice. In general, training q by sampling requires use of the log-derivative trick (aka the score-function estimator, aka REINFORCE). Eclecticos (talk) 08:37, 16 November 2025 (UTC)