CS & Physics Meet-Up by Lamarr & B3D

29 Nov 2023, 09:30 → 1 Dec 2023, 15:30 Europe/Berlin

Joseph-von-Fraunhofer Strasse 25, Floor 3, Room 302 - Lamarr Co-Working Space (Lamarr Institut, TU Dortmund University)

Ekaterina Moerova (MPI Bonn), Jens Buss (E5b), Jessica Koch (MPI Bonn), Kevin Schmidt

The "CS & Physics Meet-Up by Lamarr & B3D" aims to bring together researchers from the fields of computer science, with a focus on machine learning and artificial intelligence, and physics, specifically particle physics, astroparticle physics, and radio astronomy.

The goal of the meet-up is to provide a mutual overview of each other's research topics and research questions, as well as learning tasks, methods, and data. Our goal is to identify commonalities and develop ideas for future collaborations.

The event will include presentations, poster sessions, and open discussions to develop project ideas for future collaborations between computer scientists and physicists of Lamarr and B3D.

Wednesday 29 November

09:30 → 10:30 Arrival

09:30 Arrival & Coffee

10:30 → 11:05 Welcome & Introduction

10:30 Opening

The hosts of the Meet-Up will give a warm welcome to participants and will give a short overview of the structure of the meeting.

Speakers: Prof. Wolfgang Rhode (Physics Department, TU Dortmund University), Dr Jens Buss (Lamarr Institute, TU Dortmund University)

Opening.pptx

10:35 Introduction: Lamarr

The managing director of the Lamarr Institut, Stefan Michaelis, will give a short introduction to the institute and its research goals. The structure of the institute will be presented and an overview of the researchers will be outlined.

Speaker: Dr Stefan Michaelis (Managing Director, Lamarr Institut, TU Dortmund University)

Lamarr-Institut-Intro.pptx

10:50 Introduction: B3D

Speaker: Prof. Frank Bertoldi (Argelander-Institute for Astronomy, University of Bonn)

b3d-20231129-Dortmund.pdf

11:05 → 11:10 coffee refil

11:10 → 12:30 Overview Talks: Physics

11:10 Introduction: Radio Astronomy

20 mins Talks + 5 mins Q&A

Speaker: Dr Ancla Müller (AIRUB)

Vortrag_Lamarr_Ancla.key

Vortrag_Lamarr_Ancla.pdf

11:35 Introduction: Exp. Particle Physics

20 mins Talks + 5 mins Q&A

Speaker: Prof. Johannes Albrecht (Physics Department & Lamarr Institute, TU Dortmund University)

albrecht.pdf

albrecht.pptx

12:00 Introduction: Astroparticle Physics

Introductions to Astroparticle physics.

In particular, Gamma-ray Astronomy and Neutrino Astronomy with a glace of Simulation topics.

20 mins Talks + 5 mins Q&A

Speakers: Dr Dominik Elsaesser, Dr Tim Ruhe (TU Dortmund University)

Lamarr_Challenges_in_APP.pdf

12:30 → 13:30 Lunch

The first lunch on Wednesday will be held at the Lamarr Institute facilities to save time on the first day. During lunch, there will be an opportunity for separate exchanges between PIs and coordinators.

13:30 → 15:30 Overview Talks: Machine Learning and Computer Science

13:30 Hybrid Machine Learning for Scientific Discovery

Hybrid Machine Learning is all about inferring structured representations from empirical data, where by representations I mean transformed views of the data that make it more interpretable, or more usable for modelling and prediction. In this talk I will discuss how one can use neural networks to infer representations that satisfy partial differential equations, which one assumes model the physical processes underlying the empirical data; (ii) how simulation data from our theoretical models can be leveraged to encode mappings between infinite dimensional spaces; and (iii) how all these ideas open the door to new paradigms for scientific discovery.

Speaker: Dr Ramses J. Sanchez (Lamarr Institute, University of Bonn)

HybridML_Physics_rsanchez.pdf

13:50 Introduction to Trustworthy-AI

This presentation will give a brief overview on the various dimensions comprising Trustworthy AI in general, namely Fairness, Privacy, Autonomy, Transparency and Reliability. A specific focus is given to the last two as they are typically more concerned with the inner workings of the AI system as such. When considering AI as tool for academic, that is scientific, use understanding limitations of the systems as well as its underlying reasoning can aid the process of discovery.

Speaker: Dr Maram Akila (Lamarr Institut, Fraunhofer IAIS)

20231129_MAkila_IntroductionToTrustworthyAI.pptx

14:10 Data Bases and Data Storage

Speaker: Prof. Jens Teubner (DBIS & Lamarr Institut, Fak. Informatik, TU Dortmund)

2023-11-29_Lamarr-handout.pdf

2023-11-29_Lamarr-presentation.pdf

14:30 Coffee break

14:45 Introduction: Human-Centred Systems

Human-centered Systems are designed to interact with humans and deliver explainable and comprehensible results.

At the Lamarr Institute, we are developing human-centered approaches for bridging the gap between ML methods and human minds. On the one hand, human-centered systems adapt to human goals, concepts, values, and ways of thinking. On the other hand, these systems take advantage of the power of human perception and intelligence. Visual Analytics play a key role in combining human and machine intelligence. Thus, ML models are developed with involvement of human knowledge and then use this knowledge in generating explanations.

Speaker: Prof. Natalia Andrienko (Lamarr Institute, Fraunhofer IAIS)

HCAI_introduction_Andrienko.pptx

15:30 → 16:00 Coffee break

Time for discussions

16:00 → 17:30 Initial Zündung: Prof. Dr. Reinhard Genzel Galaxien & Schwarze Löcher (eine vierzigjährige Reise)

By chance, on the first day of the network meeting, there will be a lecture in the series "Initialzündung" of the TU Dortmund which has a thematic overlap with the network meeting. In the "Initialzündung" series, renowned scientists from all over the world who have been awarded a Nobel or Leibniz Prize, for example, are invited to TU Dortmund.

16:00 Vortrag: Galaxien & Schwarze Löcher (eine vierzigjährige Reise)

In the "Initialzündung" series, renowned scientists from all over the world who have been awarded a Nobel Prize or Leibniz Prize, for example, are invited to TU Dortmund.

Prof. Dr. Reinhard Genzel was awarded the Nobel Prize in Physics in 2020 for his research on black holes. He is director at the Max Planck Institute for Extraterrestrial Physics (MPE) in Garching and professor at the Graduate School for Physics and Astronomy at the University of California at Berkeley.

https://www.tu-dortmund.de/veranstaltungsdetail/initialzuendung-nobelpreistraeger-reinhard-genzel-an-der-tu-dortmund-35534/

Speaker: Prof. Reinhard Genzel (Max-Planck-Institut für extraterrestrische Physik (MPE) )

16:45 Q&A: Galaxien & Schwarze Löcher (eine vierzigjährige Reise)

Fragerunde nach dem Vortrag

16:01 → 17:31 Discussion groups (Open space): Networking

Open space for discussion groups

19:00 → 21:30 Social: Conference Dinner

maps_hbf.png

Thursday 30 November

09:30 → 09:50 Overview Talks

Convener: Dr Jens Buss (Lamarr Institute, TU Dortmund University)

09:30 Introduction to Embodied AI

Embodied Artificial Intelligence (AI) refers to AI that is embedded in physical systems, such as robots, and can interact with the surroundings. Embodied agents thus learn from experience in order to improve their behavior, comparable to how human learning is based on exploration and interaction with the environment. This talk will give a brief overview of the interdisciplinary field of Embodied AI and recent research activities in the context of the Lamarr Institute.

Speaker: Mr Julian Eßer (Lamarr Institute, Fraunhofer IML)

Eßer_2023_11_30_Introdcution_Embodied_AI (2).pdf

Eßer_2023_11_30_Introdcution_Embodied_AI.pptx

09:50 → 10:40 Pitch Talks: ML Methods and Application Examples (in Physics)

Convener: Dr Jens Buss (E5b, Lamarr)

09:50 ML Appl. Example Neutrino + Neutrino Map of the Universe

Speaker: Dr Mirco Huennefeld

2023_11_30_Huennefeld.pdf

10:00 Class-conditional label noise

Speaker: Dr Mirko Bunse (Lamarr Dortmund)

bunse_thu_10:00_class-conditional-label-noise.pdf

10:10 Deep Learning for real-time classification of astronomical radio signals

Continuous developments in systems for recording radio astronomical signals lead to a natural increase in the volume of data acquired. This increase, in turn, exacerbates the challenges associated with the storage and processing of this information. A potential solution to this challenge is the development of systems underpinned by deep machine learning. These systems, at the stage of signal acquisition, could classify the data and retain only those portions that contain significant scientific information for further analysis. However, the challenge resides in ensuring that these systems not only exhibit high accuracy across a vast array of signals but also maintain exceptional sensitivity to avoid overlooking weak signals that still have significant scientific value.
This presentation will discuss efforts to implement such a model, utilizing radio pulsar data from the Effelsberg telescope as a case study.

Speaker: Andrei Kazansrev (Max Planck Institute for Radio Astronomy)

pitch_talk_kazantsev.pdf

pitch_talks_kazantsev.pptx

10:20 Innovative approaches for galaxy cluster mass determination via deep learning

Determination of galaxy cluster masses from astronomical data is one of the primary prerequisites for their cosmological analysis and various deep learning methods have been tested with simulated data sets. Although the attention of the community is moving towards more advanced architectures like GANs and vision transformers, improvements can still be made with the traditional convolutional networks with simple physical inputs. These range from very simple exploitation of spherical symmetry of the objects, to training with polarization data in addition to the total intensity images. Feedback is welcome on whether there are some fundamental limitations for these convolutional networks and whether one should adopt newer models.

Speaker: Kaustuv Basu (Uni Bonn)

Basu_Kaustuv_pitch.key

Basu_Kaustuv_pitch.pdf

10:30 Representing and rendering large noisy radio data cubes

To support visual analytics of large radio/volume data, we accelerated a source finding/filtering algorithm by employing GPUs. Also, we implemented a GPU based volume renderer that achieves real-time performance and can be used in interactive environments, like an Augmented Reality setup. When visually analysing unfiltered data the frame rates of the renderer drop significantly and data is often too big to fit onto GPUs. Very first tests with a neural representation of noisy volumes to reduce storage requirements show that there is a tradeoff between quality and efficiency. We would like to learn more about DL based representation of noisy volume data.

Speaker: Prof. André Hinkenjann (Hochschule Bonn-Rhein-Sieg, Institut für Visual Computing)

Presenting Noisy Data Cubes- AndréHinkenjann-HBRS.pdf

10:40 → 11:40 Coffee break with poster session

11:40 → 12:40 Discussion groups (Open space): Open space for discussion groups

Open space for discussion groups

11:40 Deep Learning for real-time classification of astronomical radio signals

Continuous developments in systems for recording radio astronomical signals lead to a natural increase in the volume of data acquired. This increase, in turn, exacerbates the challenges associated with the storage and processing of this information. A potential solution to this challenge is the development of systems underpinned by deep machine learning. These systems, at the stage of signal acquisition, could classify the data and retain only those portions that contain significant scientific information for further analysis. However, the challenge resides in ensuring that these systems not only exhibit high accuracy across a vast array of signals but also maintain exceptional sensitivity to avoid overlooking weak signals that still have significant scientific value. This presentation will discuss efforts to implement such a model, utilizing radio pulsar data from the Effelsberg telescope as a case study.

Speaker: Andrei Kazantsev ( Max-Planck-Institut für Radioastronomie)

Zoom Room

11:40 TBA

12:40 → 14:00 Lunch

14:00 → 14:50 Pitch Talks: Inverse Problems

Convener: Dr Jens Buss (E5b, Lamarr)

14:00 Quantification

Speaker: Dr Mirko Bunse (Lamarr Institut)

bunse_thu_14:00_unfolding.pdf

Vortrag.pdf

14:10 Solving Inverse Problems with Deep Learning

Our work entails using generative models as priors for Fourier phase retrieval. We were successful in the Helsinki Tomography Challenge 2022 by employing a large synthetic dataset with end-to-end convolutional networks for limited-angle computer tomography. Additionally, our latest project focuses on extracting (ideally weak) earthquake signals from real, noisy Distributed Acoustic Sensing (DAS) data.

Speakers: Sebastian Konietzny (Informatik LS8, TU Dortmund University), Dr Tobias Uelwer (Informatik LS8, TU Dortmund University)

Solving_Inverse_Problems_with_Deep_Learning_Tobias_Uelwer_Sebastian_Konietzny.pdf

14:20 Unfolding in Astroparticle Physics

Speakers: Leonora Kardum, Prof. Wolfgang Rhode

Presentation_lamarr_Leonora_Kardum.pdf

14:30 Atmospheric Noise Removal for FYST: Current Methods and ML Prospects

The Epoch of Reionization Spectrometer (EoR-Spec) instrument on the Fred Young Submillimeter Telescope (FYST) will undertake a Line Intensity Mapping (LIM) survey targeting the [CII] line across redshifts 3.5 − 8.0. The observed frequency range for EoR-Spec, 210 to 420 GHz, is substantially influenced by atmospheric emissions that affect LIM power spectrum measurements. One of the challenges is to efficiently separate the cosmological signal from the correlated atmospheric noise. Traditional data cleaning techniques, including various filtering methods and Principal Component Analysis (PCA), are currently employed to mitigate these effects.
Machine Learning (ML) methods such as Convolutional Neural Networks (CNNs) and Gaussian Process Regression (GPR) can assist in tackling this inverse problem. In addition, incorporating outlier, glitch, and anomaly detection into the data reduction pipeline could strengthen the handling of systematics in detector timestreams.

Speaker: Ankur Dev (Uni Bonn)

AnkurDev_B3D_Lamarr_Nov2023.pdf

14:40 Leveraging Machine Learning for Next-Generation Radio Pulsar Searches

The discovery of new radio pulsars has significant implications for both Gravitational and Condensed Matter Physics. With the emergence of large datasets on the order of petabytes requiring quasi-real-time analysis, computational methods, particularly machine learning, have been increasingly important. This pitch will give a short ovreview of the progress made in applying machine learning techniques to radio pulsar searches over the last decade.

Speaker: Dr Vishnu Balakrishnan (MPIfR)

vishnu_balakrishnan_pitch_talk.key

vishnu_balakrishnan_pitch_talk.pdf

14:50 → 15:10 Overview Talks

14:50 Introduction: Resource-aware ML

The scientific area of Resource-Aware Machine Learning tries to "make the most out of a bad situation," i.e., match existing solutions' performance while only using a fraction of the resources or surpassing existing solutions' performance. To do so, we try to bridge the gap between the mathematical concepts of Machine Learning, their expression in software, and their execution in hardware. We study new algorithms for training small ML models, their post-processing (e.g., pruning) for model deployment, and their compilation to existing or new hardware.

Speaker: Dr Sebastian Buschjäger (Lamarr Institute, TU Dortmund University)

RAML.pdf

15:10 → 16:10 Coffee break with poster session

16:10 → 17:10 Discussion groups (Open space)

Open space for discussion groups

16:10 Leveraging Machine Learning for Next-Generation Radio Pulsar Searches: Challenges and Opportunities

The discovery of new radio pulsars has significant implications for both Gravitational and Condensed Matter Physics. With the emergence of large datasets on the order of petabytes requiring quasi-real-time analysis, computational methods, particularly machine learning, have been increasingly important. This session will begin with a review of the progress made in applying machine learning techniques to radio pulsar searches over the last decade. Following this, we will delve into the most recent advancements and the challenges that lie ahead. The aim is to foster a robust dialogue and establish collaborations between researchers in computer science and astronomy, working toward the next frontier in radio pulsar discovery.

Speaker: Dr Vishnu Balakrishnan (Max Planck Institute for Radio Astronomy, Bonn)

16:10 TBA

17:10 → 18:10 Wrap-Up

19:00 → 21:00 Social: Christmas Market

maps_hbf.png

Friday 1 December

09:30 → 10:30 Pitch Talks: Problems and data sources

09:30 Particle Shower Simulation

Speaker: Dominik Baack (TU Dortmund)

Lamarr.pdf

09:40 Radio Data Simulations

Speakers: Felix Geyer, Kevin Schmidt

schmidt.pdf

09:50 Potential ML applications to accelarate ISM simulation

Speaker: Tim-Eric Rathjen

2312-Dortmund-Rathjen.pptx

10:00 ML Solutions for Time Series Problems

Speaker: Dr Amal Sadallah (Lamarr Institute Dortmund)

pitch-ts-amal-saadallah.pptx

10:10 Time Series Data - Machine Learning For Observatory Sites

The CCAT observatory on Cerro Chajnantor, Chile, will be operated entirely remote - with no crew at site during observations. Hence, we require reliable Predictive Maintenance (Outlier-/Anomaly-Detection) methods on its critical infrastructure to guarantee continuous operations. Moreover, Time Series data is collected from various radiometers at Atacama representing the atmospheric precipitable water vapor. This data - supplemented by data from meteorological services - will be utilized to make 5-day forecasts in order to schedule observations. We would like to learn if any RNN model forecast supersedes that of SARIMA.

Speaker: Dr Ralf Timmermann (Uni Bonn)

LamarrePitch_Timmermann.pdf

10:20 P-adic numbers

Speaker: Prof. Thomas Liebig (Lamarrr Institut, TU Dortmund)

Talk p-adic numbers

10:30 → 11:30 Coffee break with poster session

11:30 → 12:30 Discussion groups (Open space)

Open space for discussion groups

12:30 → 13:15 Wrap-Up

13:15 → 14:15 Lunch