Valentin Thomas @_valthomas

Researcher, PhD from Mila_Quebec. Former intern at @deepmind, @Google Brain and @element_ai. Interested in all things around RL and control. valthom.github.io Joined July 2016

Tweets

97
Followers

281
Following

989
Likes

1K

Yunhao (Robin) Tang @robinphysics

4 months ago

Maybe to one's surprise, taking KL estimates as `kl_loss` to minimize does *not* enforce the KL. This implementation, however, is quite common in open source RL repos and recent research papers. In short: grad of an unbiased KL estimate is not an unbiased estimate of KL grad.

14 53 654 68K 607

Download Image

Vahid Balazadeh @vahidbalazadeh

4 months ago

Can neural networks learn to map from observational datasets directly onto causal effects? YES! Introducing CausalPFN, a foundation model trained on simulated data that learns to do in-context heterogeneous causal effect estimation, based on prior-fitted networks (PFNs). Joint…

2 10 32 3K 2

Download Image

Fabian Schaipp @FSchaipp

7 months ago

Optimization hyperparameters (LR, schedule, weight decay) do not affect loss-to-loss scaling of LLMs (which could be seen as a proxy for generalization). ☄️ Unclear: how about different optimizers (Shampoo, ScheduleFree...)? Plots from this paper: arxiv.org/pdf/2502.12120

4 9 89 6K 65

Download Image

Surya Ganguli @SuryaGanguli

8 months ago

*Every single* cure for a disease ultimately flowed from basic exploratory research. Stopping basic research is like stopping the mountain rains and expecting rivers of cures to still flow. Examples: 1) studying saliva of Gila monster -> GLP1's 2) studying funghi -> first…

Andrew D. Huberman, Ph.D. @hubermanlab

8 months ago

823 66 2K 1.2M 88

181 1K 10K 742K 1K

wh @nrehiew_

8 months ago

R1 is a ...(checks notes)... neurosymbolic architecture?

47 18 603 95K 149

Download Image

KZ @kzSlider

8 months ago

Damn, triple-homicide in one day. SAEs really taking a beating recently

10 35 369 143K 231

Download Image

Peter Battaglia @PeterWBattaglia

8 months ago

Today we’re thrilled to announce that real-time and historical AI-based weather forecasts from @Google’s WeatherNext suite of models are now available on Earth Engine and BigQuery. Anyone can access and use these data for research, analysis and operational decision making, which…

7 46 273 31K 60

Dibya Ghosh @its_dibya

8 months ago

With R1, a lot of people have been asking “how come we didn't discover this 2 years ago?” Well... 2 years ago, I spent 6 months working exactly on this (PG / PPO for math+gsm8k), but my results were nowhere as good. Here’s my take on what blocked me and what’s changed: 🧵

12 138 2K 356K 1K

Amir-massoud Farahmand @SoloGen

10 months ago

Bonjour-Hi! 1) We moved to Montreál! It is good to be back and lovely so far. 2) I joined the Department of Computer and Software Engineering of the Polytechnique Montréal @polymtl as an associate professor and Mila @Mila_Quebec as the core academic member. 🇨🇦 More news to come!

22 7 241 14K 6

Hamid Kamkari @hamid_R_kamkar

11 months ago

🚀Our tabular foundation model (TabDPT) is out and performs out-of-the-box classification/regression tasks with a single forward pass! 📊Test out the code: github.com/layer6ai-labs/… 📜Paper: arxiv.org/pdf/2410.18164… To learn more, follow the thread! 🧵(1/5)

1 3 9 526 0

Alexandre L.-Piché @alexpiche_

a year ago

Introducing ReSearch: An iterative self-reflection algorithm that enhances LLM's self-restraint abilities: • Encouraging abstention when uncertain • Producing accurate, informative content when confident Result: Significant accuracy boost for Llama2 7B Chat and Mistral 7B! 🚀

1 45 102 18K 71

Download Gif

Gergely Neu @neu_rips

a year ago

OK, time for some tweets about distances between Markov chains! Actually this is about a preprint we've just posted on arxiv with Sergio Calo, Anders Jonsson, Ludovic Schwartz & Javier Segovia-Aguas. FFO optimal transport & bisimulation. Let's dig in! arxiv.org/abs/2406.04056 1/n

8 77 406 59K 379

Download Image

Valentin Thomas @_valthomas

3 years ago

100%! Investigative papers are my favorite to write yet they are so hard to publish if you don't have SotA or an all-explaining theoretical result. However the issues they raise are sometimes what makes science progress further later on.

0 1 5 0 0

Mark Schmidt @MarkSchmidtUBC

3 years ago

@le_roux_nicolas, @BachFrancis: today it is the 10-year anniversary of the stochastic average gradient (SAG) paper getting rejected from ICML.

4 28 293 0 23

Victor Schmidt 💀🐔 @vict0rsch

3 years ago

Aaand it's a 100 ⭐️ I've been working on #PaperMemory with 1 goal: automate the recording of papers I read on @arxiv_org or @openreviewnet and the discovery code w/ @paperswithcode. It's a simple browser extension that changed my #research workflow github.com/vict0rsch/Pape…

5 30 191 0 66

Download Gif

Joseph Marino @jl_marino

4 years ago

My PhD thesis is now online (thesis.library.caltech.edu/14178/1/joseph…).

7 32 208 0 38

Marlos C. Machado @MarlosCMachado

4 years ago

We're pleased to inform you that your ICML'21 submission Beyond Variance Reduction: Understanding the True Impact of Baselines [...] has been accepted! Huge congrats to Wesley & Valentin for their persistence and hard-work! w/ @wes_chung*, @lanternol* & @le_roux_nicolas. 🧵👇🏻

Marlos C. Machado @MarlosCMachado

5 years ago

2 31 130 0 18

Download Image

2 5 55 0 3

Loopsider @Loopsidernews

5 years ago

Ça s'est passé samedi à Paris. 15 minutes de coups et d'insultes racistes. La folle scène de violences policières que nous révélons est tout simplement inouie et édifiante. Il faut la regarder jusqu'au bout pour mesurer toute l'ampleur du problème.