My PhD is coming to an end, and I'm excited to join
@IsomorphicLabs
as a Machine Learning Research Scientist soon. I'll be working in the Lausanne office starting in June. Very much looking forward to working on advancing drug discovery with AI!
normflows is a PyTorch package for normalizing flows. It can be used to model densities, even those on complex manifolds such as the cylinder surface. A brief article about the package is now available on arXiv: 1/2
Interested in doing a PhD in Machine Learning?
@MPI_IS
and
@CambridgeMLG
offer a great collaborative fellowship where you spend 1 year at one institution and 3 years at the other. Apply until December 1, 2021!
Just finished my PhD viva, being examined by
@ArnaudDoucet1
&
@pl219_Cambridge
, and passed without corrections 🥳
Many thanks to my supervisors
@jmhernandez233
&
@bschoelkopf
, co-authors, colleagues, friends, family, and everyone else who supported me throughout this journey 🙏
I'm proud to announce that I received a PhD prize by G-Research including a monetary award of £10k. Thanks
@GRESEARCHjobs
for this generous appreciation of the research I did in the past 4 years! 🎉
Today, 1:30pm CEST, I will be presenting my article Resampling Base Distributions of Normalizing Flows, being joint work with
@bschoelkopf
and
@jmhernandez233
, during a poster session at
#AISTATS2022
.
Article:
Code:
I'm delighted to announce that our work about SE(3) Equivariant Augmented Coupling Flows has been accepted as a spotlight paper at
#NeurIPS
2023 🎉 See you in New Orleans!
Our novel SE(3) Equivalent Augmented Coupling Flows can model molecular configurations based on Cartesian coordinates, incorporating equivariance and scaling well to multiple molecules while maintaining fast density evaluation and sampling.
Want to use Normalizing Flows in PyTorch? Then check out our repo .
It includes Circular Neural Spline Flows, introduced by
@DaniloJRezende
et al., that can be combined with non-periodic coordinates as well, which have not been implemented elsewhere afaik.
I'm glad to announce that our paper accompanying the normflows repository just got published in the Journal of Open Source Software 🎉
Thanks to my co-authors and all contributors to the package!
normflows is a PyTorch package for normalizing flows. It can be used to model densities, even those on complex manifolds such as the cylinder surface. A brief article about the package is now available on arXiv: 1/2
We'll present our work on SE(3) Equivariant Augmented Coupling Flows at NeurIPS next week as a spotlight poster. The poster session will be on Wednesday morning.
Also, I'm graduating soon, so feel free to reach out if you're hiring.
I'm glad to announce that our article was accepted as a spotlight presentation (notable-top-25%) at the
@iclr_conf
. Looking forward to presenting it in Rwanda!
Check out our new article Flow Annealed Importance Sampling Bootstrap (FAB), which improves training normalizing flows on target distributions without the need of samples significantly.
1/3
Our novel SE(3) Equivalent Augmented Coupling Flows can model molecular configurations based on Cartesian coordinates, incorporating equivariance and scaling well to multiple molecules while maintaining fast density evaluation and sampling.
Today, we are presenting our work "Flow Annealed Importance Sampling Bootstrap" at the AI4Science workshop
@NeurIPSConf
. Visit us in the second poster session starting at 5pm in room 388-390.
Interested in doing a PhD in Machine Learning at the
@MPI_IS
and the
@CambridgeMLG
? They offer joint fellowships where you spend time at both institutions. Apply until November 25 at
Check out our new article Flow Annealed Importance Sampling Bootstrap (FAB), which improves training normalizing flows on target distributions without the need of samples significantly.
1/3
Flow Annealed Importance Sampling Bootstrap (FAB) is a new method for fitting normalizing flows to multimodal distributions, with impressive results!
Work with amazing students and collaborators
@SilkyDogfish
,
@VStimper
,
@gncsimm
,
@bschoelkopf
. (1/8)
Interested in doing a PhD in Machine Learning?
@MPI_IS
and
@Cambridge_Uni
offer a joint position, including a fellowship with full funding. You'll spend one year at one institution and three years at the other.
Apply here:
Thanks a lot to
@TasksWithCode
for being my first-ever sponsor on GitHub! It's great that they're supporting so many contributors to open-source projects.
If you haven't seen our poster on SE(3) Equivariant Augmented Coupling Flows at the main track of NeurIPS yet, you'll have a second chance today. We'll present it at the AI for Science workshop in Hall C2, with the poster sessions being 11:10am-12:00noon and 4:35pm-5:30pm.
We will be presenting our work "AutoML Two-Sample Test" today
@NeurIPSConf
at the poster Session from 11am - 1pm in Hall J,
#536
. Looking forward to talking to you!
🧐 Do we need special-purpose libraries for powerful two-sample testing⁉️
No! All you need is an existing (Auto)ML framework and our Python package
@VStimper
@krikamol
@bschoelkopf
(and Simon Buchholz)
Just got access to
#dalle2
, getting inspiration from this art piece for my work now.
"an oil painting of a cat working on difficult scientific problems"
#dalle
#AIart
I'm attending
@NeurIPSConf
in New Orleans this week. Feel free to get in touch with me if you want to talk about Probabilistic Machine Learning, AI4Science, and more!
I played around with
#dreambooth
based on
#stablediffusion
and obtained these incredible results. Why should I even take pictures of myself anymore if I can generate them?
.
@stadtkind_rpx
,
@VStimper
,
@MariosZacharias
,
@bschoelkopf
and colleagues introduce a framework that reconstructs the energy dispersion from photoemission band mapping data and uncovers previously inaccessible momentum-space structural information.
The recording of our talk about our ICLR paper "Flow Annealed Importance Sampling Bootstrap" is now available on YouTube.
Thanks
@HannesStaerk
for being a great host!
An excellent discussion with
@VStimper
and Laurence Midgley about their "Flow Annealed Importance Sampling Bootstrap" paper - I highly recommend this one!
Definitely check it out if you are interested in Boltzmann generators.
Excited to give my first in-person talk since the start of the pandemic at the
#BiGmax
Workshop 2022 in Bochum!
Just realized that I forgot my laser pointer because I'm not used to use it anymore.
🧐 Do we need special-purpose libraries for powerful two-sample testing⁉️
No! All you need is an existing (Auto)ML framework and our Python package
@VStimper
@krikamol
@bschoelkopf
(and Simon Buchholz)
We developed a novel method to reconstruct the electronic band structure of materials using machine learning, and, finally, the article is available on arXiv!
@SilkyDogfish
and I will be presenting our work "Flow Annealed Importance Sampling Bootstrap" today at the LoG2 Reading Group. Recently, this paper was accepted at
@iclr_conf
, being among the notable top-25% of the papers.
I'm looking forward to an insightful discussion with you!
This is especially useful in the context of Boltzmann generators proposed by
@FrankNoeBerlin
,
@smnlssn
,
@jonkhler
, and Hao Wu, where we get significantly better results than previously possible when training a flow without samples on alanine dipeptide.
2/3
I'm looking forward to interesting discussions!
The paper is available on arXiv:
We also provide a 5min summary presentation on the conference website:
PhD positions in Advanced Machine Learning at Cambridge
Application deadline: noon December 3, 2020.
Details about the application process can be found here:
Interested in doing a PhD in Machine Learning?
@MPI_IS
and
@Cambridge_Uni
offer a joint position, including a fellowship with full funding. You'll spend one year at one institution and three years at the other.
Apply here:
From my experience, it is a great opportunity to work with a diverse group of bright people. Being supervised by
@jmhernandez233
and
@bschoelkopf
, I benefited from having access to the resources of both institutions and collaborating with brilliant scientists from either side.
After AI already discovered novel, more performant chip designs, new matrix multiplication algorithms are getting invented now. Recursive self-improvement seems to become a reality.
Today in
@Nature
:
#AlphaTensor
, an AI system for discovering novel, efficient, and exact algorithms for matrix multiplication - a building block of modern computations. AlphaTensor finds faster algorithms for many matrix sizes: & 1/
With the release of
#Imagen
from
@GoogleAI
yesterday, here's a quick follow-up thread on the progress of compositionality in vision-language models.🧵 1/11
@AnnalenaKofler
I have the strange feeling that I'm something in between a data physicist and a pure machine learning scientist. And once somebody invents a word for that, my role will probably have changed again anyway.
@ErikFrach
@ohhellokathrina
@masscustom
Den Terminator wird es in 5 Jahren definitiv nicht geben. Vielmehr wird man mit AI unscheinbare Prozesse effizient automatisieren, die in Summe die Welt verändern werden. Und wir sollte darauf hinarbeiten, sie zu verbessern.
Another incredibly well performing text-to-image model in such a short period of time? This cannot be real! There must be an AI generating these papers somewhere!
Excited to share our work on Parti, state-of-the-art text-to-image model that generates photorealistic images with complex compositions & rich world knowledge! Parti shows that autoregressive modeling is powerful & universally applicable! (1/n)
Blog/Paper: