Make it (net)work! - MLA2S Networking Seminar #12
12th Network Seminar of the Thematic Platform MLA²S.
IQOQI Innsbruck, Erwin Schrödinger-Saal
Zoom link: [will be published here, shortly before the meeting]
To register: click "register" below here.
-
-
1:00 PM
→
1:15 PM
News from MLA2S and Make it (net)work! 15m
Including a short round table news update.
Speakers: Dr Claudius Krause (MBI Vienna (ÖAW)), Claus Trost (Erich Schmid Institute of Materials Science of theAustrian Academy of Sciences), Elisabeth Eder (Austrian Academy of Sciences), Jan Odstrčilík (IMAFO), Kati Sparber (IGF | ÖAW), Nicki Holighaus (Acoustics Research Institute, Austrian Academy of Sciences) -
1:15 PM
→
1:45 PM
Autoregressive Transformers for Geostatistical Simulation 30m
In many spatial modelling problems, we only observe a small part of the system we want to reconstruct. For example, in subsurface modelling, boreholes provide information about geological layers or material types at only a small number of locations, while the structure between them remains unknown. The goal is therefore not to predict one unique solution, but to generate multiple realistic realizations that honor these observations but also the uncertainty in between observations. Hence, characteristic spatial patterns should be reproduced, and still show sufficient variability.
This talk will first introduce this geostatistical problem and the idea of multiple-point statistics, where new realizations are generated by matching spatial patterns against a representative training image. It will then be briefly discussed how modern generative models such as GANs and diffusion models can be applied to the task, and where their limitations lie.
The main part of the talk will focus on a Transformer-based approach to autoregressive simulation of categorical spatial fields, trained on a single representative training image. Instead of generating an entire field at once or predicting image patches, as commonly done in Vision Transformers, the model learns conditional distributions for individual locations from sparse spatial context. During training, targets and available context are varied randomly so that the Transformer learns to operate without a fixed simulation path. During generation, values are sampled sequentially and become context for subsequent predictions, allowing the same model to generate many different realizations and incorporate arbitrary conditioning observations.
The talk will also highlight several challenges that arise in this setting. A Transformer trained on a single image can exploit positional information and memorize the training image instead of learning transferable spatial structure. Conversely, restricting the model to local context improves variability but can lead to broken long-range structures, such as interrupted geological channels. Further challenges include learning from effectively only one realization and evaluating generated fields when many different outputs are equally valid and conventional pixel-wise accuracy is therefore not meaningful.
Overall, the talk explores how Transformer-based generative models can learn meaningful spatial structure from extremely limited data while balancing realism, variability, and conditioning.Speaker: Charlotte Richter (IGF | ÖAW) -
1:45 PM
→
2:00 PM
Discussion 15m
-
2:00 PM
→
2:15 PM
Break to walk a few steps and get some fresh air in. 15m
-
2:15 PM
→
2:45 PM
Medical Imaging and introduction into the new university education programme "Data, Software and AI" 30m
In many imaging problems, we only have access to incomplete or noisy measurements of the object we want to reconstruct. For example, in computed tomography (CT), images are reconstructed from projections acquired at different angles. However, acquiring sufficiently many high-quality measurements is often limited by radiation dose, acquisition time, or technical constraints. As a result, conventional reconstruction methods may produce images with substantial noise and artifacts. While deep learning has shown promising results, supervised approaches typically require high-quality reference images that are often difficult or impossible to obtain in practice.
The talk first introduces the underlying image reconstruction problem and the concept of self-supervised learning, where neural networks are trained directly from the available measurements without requiring clean reference images. Instead of learning from pairs of corrupted and ideal images, self-supervised methods exploit information within the measured data itself, for example by splitting measurements into subsets or incorporating knowledge of the underlying imaging process.
The main part focuses on how such approaches can be developed for tomographic reconstruction and applied to different imaging modalities, including X-ray CT, neutron CT, and cryo-electron tomography (cryo-ET). Particular attention is given to learning from sparse and noisy measurements, handling correlated noise, and understanding how different training strategies influence reconstruction quality. Several challenges arise in this setting, including the choice of measurement subsets, realistic detector effects, and the risk of introducing artificial structures or losing important details. These aspects become particularly relevant when transferring methods from simulated data to real-world measurements.
Beyond these research challenges, the rapid development of artificial intelligence also raises the question of how scientific advances can be transferred into practical knowledge and skills. University continuing education plays an important role in bridging this gap. At the University of Innsbruck, the new program Data, Software and AI combines programming, software engineering, data science, and machine learning in a flexible, project-oriented format. Planned uWINs, compact and research-driven intensive courses, as well as initiatives such as AI Factory Austria (AI), provide further opportunities to connect current research with practical applications and professional development.
Overall, the presentation explores how self-supervised learning can enable reliable image reconstruction from limited data, while also highlighting the broader role of universities in making advances in AI accessible beyond academic research.Speaker: Nadja Gruber (University of Innsbruck) -
2:45 PM
→
3:00 PM
Discussion 15m
-
3:00 PM
→
4:00 PM
Networking with refreshments / further discussions 1h
-
1:00 PM
→
1:15 PM