Focus on:  Hide Contributions
Back to Conference View
Login

ODSL GenAI Days Workshop

→ Europe/Berlin
Description

The Origins Data Science Lab (ODSL) is organizing a workshop on Generative AI (genAI)! GenAI models are proliferating across the natural sciences including (but not limited to):

  • Training surrogates to replace simulators
  • Fitting complicated functions
  • Density estimation (a likelihood)
  • Solving inverse problems
  • Interfacing with Simulation Based Inference
  • and many more!

The goal of our workshop is two-fold:

  1. Pedagogical introduction with hands-on examples for recent genAI models on the market (Normalizing Flows, Flow Matching, diffusion) for community members wanting to learn about new methods for research applications
  2. Provide a venue to discuss our ongoing genAI ORIGINS applications and look for synergies.

Target audience: Everyone is welcome!

We’ll be targeting the lecture / tutorial level for people who have either (1) already completed our ORIGINS data science block course or (2) already have some experience with genAI (e.g, VAEs or GANs) and are looking for what is new on the market.


Research talks: 15-20’ talks on genAI apps in ORIGINS.

  • Any contributions are welcome, also work-in-progress ones
  • Please submit by Friday, 13th Sept so we have time to adjust the agenda accordingly

**Room:** MPE seminar room (1.1.18b) 

Please register and mark your calendars!

In the registration forum, we ask some questions to understand how to tailor workshop to people who are interested in coming.

We’ll run as a hybrid event, so joining online will be available by the zoom link below:

Join Zoom Meeting

https://tum-conf.zoom-x.de/j/68019837204?pwd=bAkGvNAgfSPeGrwhbKKCPwi8F8p4ax.1

Meeting ID: 680 1983 7204

Passcode: 704536

Registration
Participants
Register
  • Monday 23 September
  • Tuesday 24 September
    • Applications
      • 4
        Integration of Augmented Reality and LLMs to Teach ALICE Detector

        Teaching complex physics topics such as the quark-gluon plasma investigated by CERN's ALICE detector is extremely important in nurturing future scientists. In this project, we aim to enhance learning by integrating the ChatGPT language model into an Augmented Reality (AR) environment that comprehensively addresses the ALICE detector.

        Within this scope, an AR application suitable for use with HoloLens2 glasses was developed, allowing students to interact with a detailed virtual model of the ALICE detector. ChatGPT was integrated into this application to enable real-time, spoken interactions through speech-to-text and text-to-speech functionalities. With this setup, a conversational learning environment was created where students can ask questions and receive personalized answers.

        A similar integration was used in an AR application developed for the quantum cryptography experiment. This application was tested in studies conducted with students. The findings show that students rated the usability of the application at an excellent level and that the feedback they received from ChatGPT improved the accuracy of their answers to the questions asked. This demonstrates that an interactive and immersive experience can make complex topics more accessible.

        The supportive role of the findings obtained from experimental applications in quantum cryptography emphasizes the effectiveness of integrating LLMs into AR for educational purposes. Applying this approach to the ALICE, reveals the potential to make complex scientific research more accessible and engaging for students.

        Speaker: Atakan Coban (Ludwig Maximilian University - Physics Education Chair)
        CNF-2024-09-24-blank-exs.pdf
        tutorial
      • 5
        Flow Annealed Importance Sampling Bootstrap meets Differentiable Particle Physics

        High-energy physics requires the generation of large numbers of simulated data
        samples from complex but analytically tractable distributions called matrix elements. Surrogate models, such as normalizing flows, are gaining popularity for this task due to their computational efficiency. We adopt an approach based on Flow Annealed importance sampling Bootstrap (FAB) that evaluates the differentiable target density during training and helps avoid the costly generation of training data in advance. We show that FAB reaches higher sampling efficiency with fewer target evaluations in comparison to other methods in high dimensions.

        Speaker: Annalena Kofler
      • 6
        Maria simulator for CMB experiments (12'+3')
        Speaker: Tony Mroczkowski (ESO)
    • 11:00
      Coffee break
    • Normalizing Flows (2/2): Flow Matching: Lecture
    • 12:30
      Lunch
    • Flow Matching: Hands On
  • Wednesday 25 September
    • Diffusion Models
      • 9
        Diffusion Models: Lecture
        Speakers: Eva Sextl (USM @ LMU), Eva Sextl (USM @ LMU)
        GENAI_FORWEBSITE.pdf
      • 10
        Closeout

        How to stay connected

        Speaker: Nicole Hartman (TUM)
        Closeout.pdf
      • 11:00
        Break
      • 11
        Diffusion Models: Tutorial
        tutorial
      • 12
        Closeout
        Speaker: Nicole Hartman (TUM)