Very excited to share our newest paper
combining machine learning & physics. We develop normalizing flows that impose the elaborate of symmetry groups you find in fundamental particle physics.
It's a beautiful mix of math, ML, and physics
10 years ago on this day I received one of the best Birthday presents one could ever hope for.
@svenkreiss
sent me a plot. We had discovered the
#HiggsBoson
!
Johann & I released v2 of our paper "Flows for simultaneous manifold learning and density estimation" with more experiments. We dubbed the class of model ℳ-Flows Here you see the flow learning the 2d manifold and the density for the Lorenz attractor! 1/n
Today I was teaching propagation of uncertainty for an introductory physics lab/data analysis class. I decided to bring automatic differentiation (via Jax) into the mix.
(the Jupyter book is in draft stage)
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
#UWMadison
alumnus
@KyleCranmer
, who played a significant role in the discovery of the Higgs boson, will be the next director of the American Family Insurance Data Science Institute. Now a physics professor at
@nyuniversity
, Cranmer starts in July 2022.
Our paper on probabilistic programming for large-scale scientific simulators is out:
pyprob is like pyro in
@PyTorch
but w/ a protocol for simulators in other languages
Awesome work
@atilimgunes
@tuananhle7
!
Big news, everyone!
@KyleCranmer
has been elected an
@APSphysics
fellow, "for the development of sophisticated statistical tools and concepts, and their application to the successful search for the Higgs boson and measurements of its properties." Congrats!
Recently
@karpathy
talked about the possibility that physics has exploits and that we should be trying to find them with RL etc. This has gotten a lot of dismissive criticism, particularly from physicists. I'd like to make a few observations / points. 1/n
RIP to Peter Higgs. The search for the Higgs boson was my primary focus for the first part of my career. He was a very humble man that contributed something immensely deep to our understanding of the universe.
An interesting and enjoyable read from Léon Bottou and Bernhardt Schölkopf. It suggests different analogies and metaphors for framing what's going on with large language models through the imagery of Jorge Luis Borge, e.g: Fiction Machines & Vindications
Happy to announce our most recent work "Flows for simultaneous manifold learning and density estimation" led by my friend and colleague Johann Brehmer
Paper:
Code:
Excited to share our newest paper combining machine learning and physics. We develop normalizing flows that respect symmetry to dramatically speed up "lattice techniques", which is a powerful computational approach to studying fundamental physics.
Today I officially left NYC. I turned in my office keys for
@nyuniversity
and my apartment keys. Packed up the fam, the dog, and hit the road.
#StillProcessing
New "A Guide to Constraining Effective Field Theories with Machine Learning"
The culmination of a 2.5 year effort in developing powerful likelihood-free inference techniques. Beyond ABC & "the Likelihood trick".
Thx JB, JP, & GL!
A call to action for the particle physics community. For 20 years we have agreed that we should publish likelihoods. We can do it technically, and recently it’s gotten better. It’s time to make this standard practice.
Our review paper on "The frontier of simulation-based inference" was published in PNAS. As always, it was great to work with Johann Brehmer and
@glouppe
. This work is gaining some traction in AI/ML for science & is relevant for public health
@PNASNews
I'm beyond excited about Monday's Machine Learning for the Physical Sciences workshop at
#NeurIPS2021
#ML4PS2021
In addition to the amazing lineup of invited speakers and panelists, we have 147 accepted papers!
While packing I found this Neural Network Primer that was given to me when I was in high school. I taught myself C and coded up my first NN with backprop in 1993. 29 years later enjoying lunch with
@ylecun
and Léon Bottou.
My
@twitter
feed has turned to crap with tons of silly viral videos and memes and hardly any of the content that I would normally see on ML / AI, data science, physics, sci-com, software, vis, etc. I mean I did have some nice memes and art in there, but it was curated before. wtf
Your attention is all I need: my talk
@CERN
today featuring RNNs, LSTM / GRU, catamorphisms, automata, GraphCNN, ProbProg, and a picture with some "French guy"
Submitted! The new chapter on Machine Learning for the Particle Data Group. The PDG provides very useful reviews on a variety of topics ranging from summaries of experimental results, theoretical models, mathematical tools, etc.
w/
@codingkazu
@seljakuros
Not long ago, I was introducing group theory with a simple example: the composition of two rotations is also a rotation. In the process, I realized that it's an easy way to remember trig sum rules, which I always forget.
(Need to know 2d rotation matrices & how to multiply them)
The saga continues… new paper using AI /ML for theoretical nuclear physics with the crew at MIT and
@DeepMind
.
“Flow-based sampling in the lattice Schwinger model at criticality”
🧵
Excited to announce a new paper with Alvaro Sanchez-Gonzalez, Victor Bapst, and
@PeterWBattaglia
(
@DeepMindAI
) on
"Hamiltonian Graph Networks with ODE Integrators"
Gives improvements in position & energy accuracy, and zero-shot generalization.
The machine learning publication model is broken. There are too few venues for publishing material that is correct that have a constructive review process. Review for big conferences is optimized to reject not to ensure correctness and improve quality.
Telescoping retweet 🧵👇
Agree with
@wellingmax
@yeewhye
@ryan_p_adams
and others on this.
Physical constraints (first volume size, then venue size) drove these arbitrary rates in the past. They are (temporarily?) gone.
Accept on interest and quality, not arbitrary targets.
New on the arXiv is a cool paper on "Machine learning action parameters in lattice quantum chromodynamics". I collaborated a bit on this project and think it is super cool. The data come from supercomputer simulations & make 4-d space-time images like this
Very happy to have this out!
"Inferring the quantum density matrix with machine learning"
together with Siavash Golkar and Duccio Pappadopulo
* Quantum Maximum Likelihood
* Quantum Variational Inference
* Quantum Flows
@NYUScience
@NYUDataScience
Mark your calendars! On Friday Dec. 15 we are having our Machine Learning and Physical Sciences (
@ML4PhyS
) workshop at
@NeurIPSConf
, and then on Saturday morning I'll be giving a talk for the
@AI_for_Science
workshop.
This guy on my flight from NYC -> SFO is trying to use a Jupyter notebook (for the first time?). He googles to get the command to open a notebook, copy/paste. That didn't work, it's not installed. Googles again... initiates anaconda download on United Wifi. Then he opens Excel 🤣
Got the
#COVIDBlues
— I could really use a break from the monotony. I love what I do, but zoom meetings and remote learning is taking its toll. I doubt I’m alone.
(And year I appreciate that I’m very privileged to have a job at all, much less one that allows for remote work)
It’s my birthday... can I get 15 followers? It doesn’t matter, but it would be funny. I tweet about physics, machine learning, math, cool images, and social justice issues
"The likelihood is dead, long live the likelihood"
An article on simulation-based (aka likelihood-free) inference in the CERN newsletter by Johann Brehmer and me.
New paper!
Hierarchical clustering in particle physics through reinforcement learning
with Johann Brehmer, Sebastian Macaluso and Duccio Pappadopulo (
@ducciolvp
)
One of the few papers using reinforcement learning for particle physics.
thread 👇
Today's the day!
Machine Learning and the Physical Sciences will be held in Room 275 - 277 at
#NeurIPS2022
. We have a great line up. First talks start at 8am!
Please use hashtag
#MP4PS2022
Important Update: We have extended the submission deadline for the Machine Learning and Physical Sciences workshop at
#NeurIPS2022
to September 29
#ml4ps2022
Several of us felt the need to prepare a rebuttal to the offensive & unsound talk at CERN last week and reaffirm our values. I’m very happy that over 200 physicists have already signed on.
New 500 page book using techniques from theoretical physics (effective field theory) to understand deep neural networks.
@danintheory
@ShoYaida
@BorisHanin
Blog post 👇
In our newest paper we discuss the frontier of simulation-based inference (aka likelihood-free inference) for a broad audience. We identify three main forces driving the frontier including:
#ML
, active learning, and integration of autodiff and probprog.
I’m saddened by the developments at google around the firing of
@timnitGebru
. I don’t know enough about what happened to comment, but the response to the situation in the last few hours definitely shows AI/ML has some problems as a community in terms of respect and inclusion.
Our review of Machine Learning and the Physical Sciences made the cover of
@APSphysics
Review of Modern Physics !
(or here )
I worked on that image for a while and it incorporates the famous CNN figure from
@ylecun
BREAKING NEWS
The Royal Swedish Academy of Sciences has decided to award the
#NobelPrize
in Physics 2018 “for groundbreaking inventions in the field of laser physics” with one half to Arthur Ashkin and the other half jointly to Gérard Mourou and Donna Strickland.
This is cool! Instead of generic SGD, optimization layers represent the optimization of convex optimization problems directly: and the solution is itself differentiable.
@brandondamos
New paper, it's a beast! Huge effort led by Dan Hackett. We investigate different ways of training flows (reverse & forwards KL) for multi-modal distributions (eg. scalar field theory), combining them with MCMC samplers, & pros/cons of performance metrics
Dear physics / astro peeps:
Please let me know your favorite text books (or similar resources) for probability and statistics at the graduate level. I'm looking for something that is somewhat pedagogical and won't seem inaccessible to physicists with no prob/stats background
Wow, even more impressed by Grant Sanderson
@3blue1brown
.
Insightful thoughts on engagement and education, eloquent, and makes me think about “the unreasonable effectiveness of mathematics” from a different angle than application to physics
Did you know that Sau Lan Wu (my PhD advisor)
@UWMadPhysics
was a key figure in the discovery of three fundamental particles: the charm quark, the gluon, and the Higgs boson? Celebrating
@UWMadison
's 175th birthday
#UW175
Wacky article arguing that instead of calculus, we should teach statistics and computer science to high school students.
Statistics Lesson 1: ∫ p(x) dx = 1 ... what's an integral?
Machine Learning Lesson 1: ∇_θ L[y,f(x, θ)] ... what's a derivative?
I'm giving a talk tomorrow for ML4Jets. I noticed that
@mmbronstein
's Deep Learning zoo didn't have a Recurrent / TreeRNN, so I added one. Note the legs 😂
We've updated our the website for Machine Learning in the Physical Sciences with the schedule and more than 150 contributed papers! Many thanks to
@glouppe
and
@atilimgunes
for the website updates, all the organizers and all the reviewers!
#ML4PS2020
Happy to announce our most recent work "Flows for simultaneous manifold learning and density estimation" led by my friend and colleague Johann Brehmer
Paper:
Code:
@GaryMarcus
@frossi_t
Fantastic people, but do you think that the presence of big names should guarantee acceptance? That sounds like privilege to me. For our workshop we put in a lot of thought about format, accessibility, diversity, value to the community, etc.
Was just thinking about differentiable sorting... google... may 2019... dang!
Differentiable Sorting using Optimal Transport:The Sinkhorn CDF and Quantile Operator
Marco Cuturi, Olivier Teboul, Jean-Philippe Vert