In approaches using deep projective priors, we link a key geometrical attribute, the "orthogonality of the projection" with identifiability and convergence rate.
Using this attribute to regularize the learning of such priors improves stability and robustness for ill posed imaging problems.

Pour être tout à fait honnête, "se dessine avec une règle" est dans la leçon. Je ne sais pas si la règle utilise l'axiome du choix par contre

a good basis to understand what is at play when we try to improve PTQ, by e.g. cross layer equalization or adaptive quantization.

However, it is always possible to try to study algorihms as tools to minimize a recovery error to try to bypass this 2-step process (which I have tried to do lately).

I wonder if such approach is possible for the pure learning problem.

So it makes sense to study optimization of the loss.

There is a parrallel in inverse problems, set up a function to minimize, guarantee convergence AND that minimizers identify the righ objects (that last part being often overlooked).
↓

Not sure if we are talking about the same thing, I was thinking about that:

Zoran, D., & Weiss, Y. (2011, November). From learning models of natural image patches to whole image restoration, ICCV

I used it for estimation of low rank GMM from compressed patch database:
hal.science/hal-03429102

but natural image patches do (did not read the article though)

Rencontre incontournable de la communauté ! Je n'y serais pas personnellement mais Ali Joundi y présentera nos travaux sur les autoencodeurs atomiques

cnrs.hal.science/hal-04773954/

My workaround without a bib file: for a personal ref section : [YT1], [YT2]...

\newcounter{bibc}
\newcommand\nbib{\arabic{bibc}}
\stepcounter{bibc}\bibitem[YT\nbib]{label1} Reference 1
\stepcounter{bibc}\bibitem[YT\nbib]{label2} Reference 2

I dont remember why I couldn't do it with multibib

I like to think about them as Schrödinger papers, both accepted and rejected at the same time. 🤣