Nicola Branchini

The true academic method: overthink, underdeliver, cite yourself.

24. arxiv.org/abs/2510.00389
'Zero variance self-normalized importance sampling via estimating equations'
- Art B. Owen

Even with optimal proposals, achieving zero variance with SNIS-type estimators requires some innovative thinking. This work explains how an optimisation formulation can apply.

Autodifferentiable Ensemble Kalman Filters | SIAM Journal on Mathematics of Data Science Data assimilation is concerned with sequentially estimating a temporally evolving state. This task, which arises in a wide range of scientific and engineering applications, is particularly challenging...

were you reading:
epubs.siam.org/doi/abs/10.1...

ELLIS PhD Program: Call for Applications 2025 The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...

I'm looking for a doctoral student with Bayesian background to work on Bayesian workflow and cross-validation (see my publication list users.aalto.fi/~ave/publica... for my recent work) at Aalto University.

Apply through the ELLIS PhD program (dl October 31) ellis.eu/news/ellis-p...

"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...

It’s a JAX, JAX, JAX, JAX World
statmodeling.stat.columbia.edu/2025/10/03/i...

I am happy to announce that the Workshop on Emerging Trends in Automatic Control will take place at Aalto University on Sept 26.

Speakers include Lihua Xie, Karl H. Johansson, Jonathan How, Andrea Serrani, Carolyn L. Beck, and others.

#ControlTheory #AutomaticControl #AaltoUniversity #IEEE

Just finished delivering a course on 'Robust and scalable simulation-based inference (SBI)' at Greek Stochastics. This covered an introduction to SBI, open challenges, and some recent contributions from my own group.

The slides are now available here: fxbriol.github.io/pdfs/slides-....

YouTube video by Team Cherry Hollow Knight: Silksong - Special Announcement

The countdown is on!

Join us in 48 hours for a special announcement about Hollow Knight: Silksong!

Premiering here: youtu.be/6XGeJwsUP9c

YouTube video by International Centre for Theoretical Sciences The Mathematics of Large Machine Learning Models (Lecture 1) by Andrea Montanari

Turing Lectures at ICTS
www.youtube.com/watch?v=_fF6...
www.youtube.com/watch?v=mGuK...
www.youtube.com/watch?v=yRDa...

"Io stimo più il trovar un vero, benché di cosa leggiera, che ‘l disputar lungamente delle massime questioni senza conseguir verità nissuna"

Today I learnt this Galileo Galilei quote:

"I value more the finding of a truth, even if about something trivial, than the long disputing of the greatest questions without attaining any truth at all"

Feels like we could use some of that in research tbh..

It is somewhat amusing to see other reviewers confidently and insistingly rejecting alternative proposals (in suitable settings) to SGD/Adam in VI/divergence minimization problems

algorithmsbook.com/validation/f...

Flying today towards Chicago 🌆 for MCM 2025

fjhickernell.github.io/mcm2025/prog...

Will give a talk on our recent/ongoing works on self-normalized importance sampling, including learning a proposal with MCMC and ratio diagnostics.

www.branchini.fun/pubs

Really cool work : ) @alexxthiery.bsky.social

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

agree; you should check out @yfelekis.bsky.social 's work on this line 😄

Just don't see that the PPD_q of the original post leads somewhere useful.
Anyway, thanks for engaging @alexlew.bsky.social : )

I agree, except I think it can be ok to shift the criteria of "good q" to instead some well-defined measure of predictive performance (under no model misspecification, let's say). Ofc Bayesian LOO-CV is one. We could discuss to use other quantities, and how to estimate them, ofc.

Genuine question: what is the estimated value used for then ?

(computed with the inconsistent method)

Well, re: [choose q1 or q2 based on whether P_q1 > P_q2]
I seem to understand that many VI papers say: here's a new VI method, it produces q1; old VI method gives q2. q1 is better than q2 because test-log PPD is higher !

Not entirely obvious to me, but I see the intuition !

Am definitely at least *trying* to think carefully about the evaluation here 😅 😇

Right ! Definitely not sure if necessary, but I like to think there would be value / would be interesting if we wanted to somehow speak formally about generalizing over unseen test points

It still seems "dangerous" to use the numerical value of (an estimate of) ∫ p(y|θ) q(θ) to decide which approximate q is better.

(Of course, you may argue we maybe shouldn't use even any MC estimates of the original ∫ p(y|θ) p(θ|D) with q as proposal, but the above is even less justified)

I don't see that it needs to get that philosophical ?
It is totally possible to formally estimate the pdf itself, since we have some 'test' samples of y, and consider MISE type errors, even if in this case pointwise evaluations of the pdfs have the intractable integral.