Very excited to finally share the excellent work of my PhD students
@RochmanOmer
and
@SachaLewin
on NeuralMPM, a neural emulation framework for particle-based simulations!
In Belgium, scientific papers that originate from public funds can now all be made public in open access, regardless of any contract with publishers. It is written in the law. And retroactive.
Voilà, la loi fédérale belge autorisant
#OAGreen
publiée aujourd'hui au Moniteur belge art. 29 page 68691
TOUS articles scientif des chercheurs belges peuvent désormais être déposés en
#OA
sur
#RI
avec embargo max 6/12 mois. Quel que soit contrat signé !
Excited to kick off the second term with my students! Time to teach Deep Learning again! This year, I have decided to say goodbye to RNNs and GANs and cover GNNs and score-based diffusion models instead 🤖💡
Bayesian nugget of the day: Metropolis-Hastings. 1) Draw a candidate θ from a proposal q(θ|θₜ), 2) Accept θ if u < π(θ)/π(θₜ) q(θₜ|θ)/q(θ|θₜ) for u~U[0,1], 3) Repeat. Beautifully, this results in a Markov chain whose stationary distribution coincides with the target π!
Finally done with teaching my Deep Learning class for this term! This year's version includes new lectures on autodiff and transformers. All (787) slides can be found at 🧑🎓
Scientific and societal impact of research comes rarely from where we expect. Open source software is one such example: today, Scikit-Learn just passed 10000 citations! Congratulations to all contributors for all the progress they triggered!
Remember the Galton board? When nails are positioned such that the probability of bouncing left is always the same, the resulting distribution is a binomial, for which the likelihood p(x) of bin x is known analytically.
Today I used
@OpenAI
Davinci model (GPT-3) for an actual Turing test in my first lecture of Introduction to AI. It turned out great! Davinci passed the test! 🤖
Tomorrow I'll actually teach my first ever lecture on deep learning, rebuilding all the basic constructs, starting from LDA up to deep rectified networks. Lecture slides available at
Watching PhD students lose their last crumbs of illusion about science when faced with stubborn, unreasonable, or completely absent reviewers is heartbreaking. We've got to do better as a community.
#NeurIPS2024
Started training my baby, hand-made 40M GPT model on Harry Potter books 🤓 GPUs go brrrr! 🚀 (
@PyTorch
's nn.DataParallel makes multi-GPU support so easy <3)
New paper by my student
@WehenkelAntoine
: Unconstrained Monotonic Neural Networks! In this work, we show how to make use of free-form neural networks to model arbitrary monotonic functions and derive a new autoregressive flow.
Brilliant step-by-step review of optimization methods for large-scale machine learning, by Léon Bottou et al. Must read if you want to understand why SGD is actually a good optimization algorithm in ML!
The Kalman Filter, as implemented in assembler in the Apollo Guidance Computer! Those little things you discover when preparing your next lecture on temporal models...
Who knew? Changing the reset gate in
@kchonyc
's GRU so that the gate can take values in ]0,2[ instead of ]0,1[ makes the cell bistable, hence enabling long-lasting memories! Very nice work by my colleagues
@vecovennicolas
,
@DamienERNST1
and Guillaume Drion!
We have taken inspiration from biological neuron bistability to embed Recurrent Neural Nets (RNN) with long-lasting memory at the cellular level.
Excellent performances on time-series which require very long memory.
#DeepLearning
#ArticialIntelligence
The more I talk about simulation-based inference, the more I realize that the concept of an intractable likelihood is completely foreign in some fields. A short thread 🧵
Had quite some fun this morning teaching a new lecture on
#GNNs
! 🧑🏫 Key takeaway: GNNs are built on the blueprint of stacking shared, local, and permutation invariant functions. They generalize classic architectures like convnets & transformers to graph-structured data 🤯🌐
📢 My research group has an open position for a PhD candidate in deep-learning for simulation-based inference. Details available at PM for further details 🤖🔭
You say Normalizing Flows? We see Bayesian Networks! Quite excited to announce two papers
@WehenkelAntoine
and I have just released on arXiv! + Thread below 👇
I am VERY excited to share not 1 but 2 papers hot from arxiv, joint work with
@glouppe
!🥳
We are showing strong connections between Bayesian networks and normalizing flows.
1)Using Bayesian networks to better understand NFs and 2)increasing BN framework with the NF machinery.
Bayesian recipe of the day 🧑🍳: Take a simple latent variable model where the latents are Gaussian and the observed variables are linear Gaussian. Fit the linear projection parameters by maximum likelihood estimation. Then, posterior inference gives you (Probabilistic) PCA! 🤯
Finally found some time for hacking around with JAX! It feels almost as if I was back in 2011, writing Numpy code for sklearn, except now it is fully differentiable and runs on GPUs 🤓
Variable importances computed from random forests are Shapley values! Happy to share an 8-year follow-up to my very first
#NeurIPS
paper, back in 2013. [accepted at
#NeurIPS2021
] 🧵
Diffusion models are not only good at generating pretty pictures, they can also solve (very) high-dimensional inference problems! Really proud of this last piece of work led by
@FrancoisRozet
on Bayesian data assimilation 🌍🤖 (We scale to 128x64x64-dimensional posteriors!)
📢 Tired of inaccurate weather forecasts? Of course you are, and that's why
@glouppe
and I are thrilled to announce Score-based Data Assimilation, a novel trajectory inference method powered by score-based generative models. 🧵
Today is the first lecture of my new Deep Learning course at
@UniversiteLiege
. We will start easy with some reminders on the fundamentals of machine learning. Materials will be posted every week at Feedback is welcome!
The list of all 191 papers accepted at the
#ML4PS2023
NeurIPS workshop is now available at Congratulations to all authors and huge thanks to all reviewers! This year, we reached an unprecedented number of 3.9 reviews per paper on average!
Phew! I am now halfway through my deep learning lectures. Today we went over recurrent networks, starting from Elman networks up to Differentiable Neural Computers. Slides available et
At the Isaac Newton institute for mathematics for a few days. Where you find black boards in the bathroom. You cannot be too cautious in case of emergency.
Lazy tweet: I am looking for applications of automatic differentiation beyond neural nets. Any exciting example I could show to my students? Thinking about things like differentiating through simulations or gradient-based hyper-parameter optimization
In previous work, we showed that algorithms for simulation-based inference can all produce underdispersed posteriors, which can make them unreliable for scientific inquiry. Today, we are happy to announce a first algorithm for reliable simulation-based inference! 🧵
Great news, our Machine Learning and the Physical Sciences workshop at
#NeurIPS2022
was accepted!
This will be the fifth edition of this fantastic workshop series.
More details to follow.
#ML4PS2022
This year, I am really satisfied with the second version of my Introduction to AI lectures. Tomorrow will already be the last lecture! All (700+!) slides can be downloaded at
Happy to announce our latest work led by
@ArnaudDelaunoy
for fast posterior inference on gravitational wave data using likelihood-to-evidence ratio estimation (AALR / SNRE), reducing inference time from days to minutes or less! 🔭🤖
Digging into old papers to prepare a deep learning talk for a general audience. I never realized the neocognitron paper also came with an attention model!
Look what I found in the office! Yesterday, I gave my AI class on top of the very first two volumes of the
#NeurIPS
proceedings (1987, 1988). Stand on the shoulders of giants they say... 🤖
The first-ever image of a black hole (M87 or Virgo A, left) is here compared with a simulation (middle) and the simulation blurred to the expected resolution of the telescope (right)
Tomorrow is already my last Deep Learning lecture of the year! The full deck of (600+) slides can be downloaded at
Quite happy with the final result, but relieved at the same time! Now let's hope for the best regarding the outcome of the exam.
📢 Open PhD position in my group, on simulation-based inference and generative models for particle physics, on a joint project with Fabio Maltoni and Tilman Plehn
Big news! 🚨 Our workshop got accepted at ICML 2023! Join us to discuss how to combine scientific ⚗️ and ML 🤖 models into hybrid models 🧌. Let's bridge the gap between fields! 🤝 Stay tuned for more details 📻
#ICML2023
Can we run reliable simulation-based inference on a budget? when having as few as O(10-100) simulations? Yes, we can! Check out
@ArnaudDelaunoy
's latest paper on Bayesian neural networks for low-budget simulation-based inference 🤓
What if you want to do posterior inference but don't have a prior to start with? In this new work led by
@VandegarM
, we investigate how the empirical Bayesian can use neural density estimators to estimate a source distribution over uncorrupted samples from noise-corrupted ones.
Open PhD position
@UniversiteLiege
in exoplanet imaging with deep learning. Exciting opportunity to develop ML algorithms for the detection of new planets!
Very proud to sign my first paper with
@ATLASexperiment
on "Deep Generative Models for Fast Photon Shower Simulation in ATLAS"! This project actually began in 2017, way before generative AI was cool. It is a relief to see it finally published 🤖
In our latest paper with Johann Brehmer,
@KyleCranmer
and
@juanpavez
, we compare this extraction process to 'mining gold' from implicit models and show how it can dramatically improve data efficiency of likelihood-free inference algorithms.
Simulation-based inference outside of science? Yes! Here is our attempt for robotic multi-fingered grasping 🤖 It took some efforts to bridge the sim-to-reality gap, but adding nuisance parameters was then natural and worked very smoothly.
Less talking, more action: tomorrow in my deep learning class, I'll be building and training a mini GPT from scratch👨💻! It won't rival the largest LLMs 🦙but expect some magical Harry Potter-esque babble✨🪄
Quite satisfactory to reach a point in my AI class where students should have all the pieces for building an approximately intelligent agent in partially observable stochastic environments, combining algorithms for reasoning over time, handling uncertainty and taking decisions.
TIL when preparing my second deep learning lecture that the operators' manual for Rosenblatt's Mark 1 Perceptron machine was a classified document. It became unclassified only in 1977. The manual can now be found at
Mission complete! It took us several hours (almost 4!), but we built and trained a fully-working GPT language model on Harry Potter! The code implemented in class can be found at
Less talking, more action: tomorrow in my deep learning class, I'll be building and training a mini GPT from scratch👨💻! It won't rival the largest LLMs 🦙but expect some magical Harry Potter-esque babble✨🪄
This is the model from Report 13 of the Imperial College COVID-19 response team. It fits the death data jointly from 11 European countries to estimate the reproduction number and the effect of lockdowns. Such a remarkable piece of
@mcmc_stan
Looking for a PhD studentship? Want to develop AI and deep learning for furthering scientific discovery? Let's get in touch! More details at I'll be at
#NIPS2017
if you want to chat.
Inspired by , today I tried to distill to my students some intuition behind the effectiveness of deep learning by folding a sheet of paper in front of them. Not sure if they clicked, or if I just lost them...
The call for papers to our Machine Learning and the Physical Sciences workshop
#ML4PS2021
#NeurIPS2021
is now out! The deadline is set a month from now, on September 18.
What if you want to do posterior inference but don't have a prior to start with? In this new work led by
@VandegarM
, we investigate how the empirical Bayesian can use neural density estimators to estimate a source distribution over uncorrupted samples from noise-corrupted ones.
Excited to have this finally out! In this work, we look to see how a neural network can hijack random number calls in a particle physics simulator for amortized posterior inference. Praises to
@atilimgunes
and
@lukasheinrich_
!
Exciting news for
#AI
research in Belgium! Fine print: I will soon be looking for new PhD students to work on AI for science! Prospective applicants are welcome to reach out 🤖🧑🎓
In this first part, I illustrate the change of variables theorem, together with its application for parameter estimation of affine transformations. Purely artificial I concur, but that's nice exercise for getting used to JAX API.
Stay tuned for Part 2!
2nd paper of the day with friends from
@GRAPPAinstitute
, incl.
@C_Weniger
,
@NilanjanBanik
and
@gfbertone
! We haven't solved protein folding, but I am quite excited about this study in which we show the full potential of simulation-based inference for grand scientific problems💫🤖
New paper today w/
@joeri_hermans
, Nil Banik,
@C_Weniger
and
@glouppe
: a likelihood-free Bayesian inference pipeline to infer properties of dark matter from observations of stellar streams
Really impressed by the expressiveness of the latent space of
#StyleGAN
! I can map myself into latent space and interpolate to my office neighbor
@DamienERNST1
Brilliant conversation between
@EtienneKlein
and
@ylecun
: can intelligence be artificial? (in French) Feels refreshing to see a debate of this quality among all the over-hyped AI coverage
The third edition of my "Introduction to AI" class -- a modernized summary of GOFAI -- is now finished! Now is time to relax during the Christmas break 😴 Slides are all available at .
Hot from your arXiv: a (much) revised version of from my student
@joeri_hermans
, where we show how to effectively plug approximate likelihood ratios to obtain likelihood-free variants of MCMC algorithms.
Meh, rejected from
#AISTATS2022
because the "paper focuses too much on applications in astronomy"?! An argument that was not even raised in any of the original reviews... Alright then, we will get a ticket for the next lottery.🎲
First paper of the year, by my student
@joeri_hermans
! In , we adapt MCMC algorithms to the likelihood-free scenario by making use of a parameterized likelihood ratio model learned by supervised learning.
👇 Proud of my student who reimplemented almost entirely from scratch a full deep learning pipeline for voice embedding, cloning and synthesis. Let him know what you think!
The finished product of my master's thesis: a real-time voice cloning toolbox. Clone a voice from only 5 seconds of audio and use it for text-to-speech!
#DeepLearning
#AI