Anyone at #univie or #TUW looking for a Master's #Thesis at the intersection of particle physics and machine learning? Let me know! .#HEPML #HEPHY.

I heard that the sky has a better color in other parts of the internet: .Go and find me there under the same handle. #eXit.

RT @GregorKasieczka: New(ish) result: A massive comparison of generative architectures as surrogate models for particle showers in calorime….

RT @GregorKasieczka: Today ends the #ML4Jets conference in Paris. Great to see the progress in the field from one year to the next, with n….

Congratulations, @MArcelaCArena !.

Now @RPWinterhalder gives an overview on ML-assisted event generation, including MadNIS!

Good Morning from #ML4Jets in #Paris! Looking forward to many great presentations and discussions on #ML for #HEP

Thanks to everyone who contributed to this challenge with hard work, discussions, submissions, comments, insights, and more! .I'm now looking forward to the next generation of smart fast simulation for particle physics. 12/12.

More plots (and tables) are at .Data and notebook used to create all figures will soon be at too. #OpenData #FAIR. 11/12.

Or:.How the binary AUC aligns with the multiclass log-posterior. 10/12

For example:.Training a binary classifier on low-level (voxel) features vs. on high-level (physics observables) features. 9/12

But wait, there is more! . With such a large dataset of submitted samples we can also study how different quality metrics correlate, which is an important question well beyond calorimeter simulation and high-energy physics!. #evaluation #generativeAI. 8/12.

. and ds2+3. The overall story consistent across all datasets: .- Diffusion/CFM models have best quality, but are slow in sampling.- VAEs/GANs have worse quality, but are fast.- Normalizing Flows have a reasonable trade-off, but don't scale to dataset 3. 7/12

The main results are the Pareto fronts in "quality" vs "time". As representative, we picked the multiclass log-posterior and the GPU batchsize 100 as metrics. Here are the #ParetoFronts for ds1. 6/12

We are now able to compare all of them in terms of "generation speed" (on CPU/GPU + diff batchsizes), "resource requirements" (num of parameters), and "sample quality". These include: binary AUCs, multiclass log-posteriors, KPD/FPD, and precision/recall/density/coverage. 5/12

We got 59 submissions based on 31 different generative models from 23 different collaborations. It's nicely spread across all 4 datasets and generative architectures: #GAN s, #VAE s, #NormalizingFlows, #DiffusionModels, and models based on Conditional Flow Matching (#CFM).4/12

All datasets have the same general format: Showers are discretized to voxels in a cylindrical coordinate system, defined by the incident particle hitting the calorimeter surface. Datasets are: .ds1 - photons: 368 voxels.ds1 - pions: 533 voxels.ds2: 6480 voxels.ds3: 40500 voxels

More than 2 years ago, we started a community challenge by publishing 4 calorimeter shower datasets & asking the community to train their favorite generative AI model on the conditional probability p(shower | E_incident). More details at 2/12.

I'm happy and proud to present the results of the #CaloChallenge! . A huge thank you to everyone who contributed to this effort in the past year(s)!.1/12

Join us next Thursday for #DarkMatterDay and learn how #MachineLearning can assist the search, too!.

RT @angryfermion: it's not machine learning unless it's from the machine region of france. otherwise it's just sparkling linear algebra.