ELLIS Unit Linz: Pioneering Efficient, Adaptive, and Scientifically Grounded Deep Learning
Mission & Vision
The ELLIS Unit Linz envisions advancing AI as a transformative technology by pioneering deep learning approaches that are adaptive, efficient, robust, and scientifically grounded. The Unit aims to be a global leader at the intersection of fundamental AI research and real-world impact—from core architectures and algorithms to applications in life sciences, physics, music, and industrial settings—empowering society with intelligent systems that understand, learn, and act with purpose.
The Unit pursues this vision along four strategic pillars:
Cutting-Edge Interdisciplinary Research: Developing novel methods in efficient sequence modeling, foundation models, uncertainty quantification, and neuro-symbolic AI to solve complex, real-world challenges across scientific and industrial domains.
Fostering Young Talent and Scientific Excellence: Cultivating the next generation of AI leaders through PhD programs, postdoctoral research, and tenure-track opportunities, with a focus on hands-on, high-impact projects. The Unit strives for international recognition through top-tier publications, keynotes, and open benchmarks.
Bridging Research and Real-World Impact: Translating AI breakthroughs into practical tools—such as MHNfs for drug discovery and PaSST for audio analysis—and industry collaborations addressing pressing societal challenges. Public engagement through initiatives like Ars Electronica exhibitions and the Kepler Awards helps democratize AI innovation.
Strengthening Global and National AI Ecosystems: Leading Austria's AI excellence as the central hub of the FWF Excellence Cluster BILAI (Bilateral AI), which integrates symbolic and subsymbolic AI into next-generation neuro-symbolic solutions. Collaborating with international partners across the ELLIS network and beyond to shape the future of AI on a global scale.
Key Numbers & Achievements
Selected as one of the initial 17 ELLIS Units in December 2019, the ELLIS Unit Linz is a founding member of the ELLIS network and a leading center for AI research in Austria. The Unit is hosted at Johannes Kepler University Linz (JKU) and builds on the LIT AI Lab, a permanent research center established in 2017 with three founding groups: Deep Learning (Sepp Hochreiter), Logical Reasoning (Armin Biere), and Computational Perception (Gerhard Widmer). Joining ELLIS was a natural step to embed JKU's AI expertise into a pan-European network dedicated to keeping Europe competitive in modern AI.
Since its founding, the ELLIS Unit Linz has grown into a vibrant interdisciplinary research ecosystem comprising specialized groups in Deep Learning, AI in Life Sciences, Computational Perception, Human-Centered AI, Symbolic AI, and Computational Data Analytics, among others. The Unit currently comprises 12 members—3 Fellows, 2 Scholars, and ELLIS 7 Members—and supports roughly 30 early-career researchers across its research groups, with further growth planned through the ELLIS PhD Program.
Key milestones include:
Falling Walls Breakthrough of the Year 2021 (Category Arts & Science) for an AI-based, expressivity-aware real-time music co-performer, the ACCompanion (IJCAI 2023)
NeurIPS Posner Lecture (2024): Sepp Hochreiter delivered the prestigious Posner Lecture under the title "Toward Industrial Artificial Intelligence" at the world's leading AI conference
Wilhelm Exner Medal awarded to Sepp Hochreiter for fundamental contributions to artificial intelligence, in particular Long Short-Term Memory (LSTM) networks
Election of Gerhard Widmer and Sepp Hochreiter to the Austrian Academy of Sciences (ÖAW)
FWF BILAI Excellence Cluster (2024–2029): JKU leads Austria's nationwide effort to integrate symbolic and subsymbolic AI into next-generation neuro-symbolic solutions
Launch of spin-offs NXAI and Emmi AI, translating fundamental research into industrial applications and demonstrating the Unit's commitment to innovation beyond the lab
Awards:
ECML-PKDD 2024 Test of Time Award for the most influential paper published at ECML-PKDD 2014 ("Large-Scale Multi-label Text Classification – Revisiting Neural Networks")
Best Student Paper Award at ECML-PKDD 2024 for work on explainable prototype-based recommendation systems
Best Paper Award at the ELLIS Machine Learning for Molecules Workshop for Bio-xLSTM
Best Student Paper Award at IEEE Computational Games (IEEE CoG) 2024
Best Paper Award at IEEE Computational Games (IEEE CoG) 2021
Research Areas & Highlights
ELLIS Unit Linz conducts both foundational and applied AI research across several key domains:
Core AI and Deep Learning Architectures
At the heart of the Unit's research is the development of novel deep learning architectures and efficient sequence models. The xLSTM architecture extends the foundational LSTM framework with exponential gating and new memory structures, achieving competitive performance with transformer-based models while offering improved efficiency. The TiRex foundation model for time series brings large-scale pre-trained modeling capabilities to temporal data across diverse domains. Research on uncertainty estimation—including mixture density networks, Monte Carlo dropout, and conformal prediction—provides critical tools for reliable AI deployment in high-stakes applications such as hydrology, drug discovery, and clinical decision-making.
AI for Drug Discovery and Life Sciences
The Unit develops machine learning methods for molecular and biomedical applications. MHNfs (Modern Hopfield Network for Few-Shot Bioactivity Prediction) enables activity prediction with minimal labeled data, validated on PubChem bioassays and demonstrating real-world applicability ([Journal of Chemical Information and Modeling, 2024](https://doi.org/10.1021/acs.jcim.4c02373 )).
VN-EGNN introduces equivariant graph neural networks with virtual nodes for enhanced protein binding site identification ([Journal of Cheminformatics, 2025](https://doi.org/10.1186/s13321-025-01127-9 )).
Further contributions include CLOOME for bioimaging and ConGLUDE for molecular representation learning.
AI for Physics and Earth System Simulation
The Unit advances AI-driven scientific simulation through physics-informed deep learning. The Aurora foundation model for Earth systems, trained on over one million hours of geophysical data, outperforms operational forecasts for weather, air quality, ocean waves, and tropical cyclones at orders-of-magnitude lower computational cost, with fine-tuning capabilities for diverse applications including climate modeling ([Nature, 2025](https://doi.org/10.1038/s41586-025-09005-y )).
Neural operator surrogates based on Fourier Neural Operators (FNOs) accelerate plasma simulations, with transfer learning reducing errors for fusion research applications ([Nuclear Fusion, 2025](https://doi.org/10.1088/1741-4326/adfdfb )).
Computational Perception: Audio and Music AI
JKU's Institute of Computational Perception has become one of the world's leading research labs in acoustic scene understanding and music AI, consistently winning top ranks in DCASE challenges for tasks such as low-complexity acoustic scene classification and language-based audio retrieval. The transformer-based PaSST model has become a standard embedding model used in Fréchet and KL Audio Distances for evaluating large audio generation systems such as MusicGen and Stable Audio. In music information research—recognized by two ERC Advanced Grants—tools such as Madmom and BeatThis! are routinely used in commercial digital music applications worldwide. Current research focuses on information-theoretic aspects of music perception, including the development of a Musical Expectation Benchmark Suite, a major international multi-disciplinary effort. The Unit also pursues AI-based computational musicology, including the launch of the HINVES Spirio Prize, a novel piano competition designed to advance research on AI models of musical expressivity.
Human-Centered and Trustworthy AI
The Human-Centered AI group develops methods for debiasing, privacy protection, and explainability in recommendation systems. Novel techniques "unlearn" harmful biases from latent user representations during training, leading to less stereotypical recommendations while simultaneously protecting user privacy ([ACM, 2024](https://dl.acm.org/doi/10.1145/3705328.3759320 ); [Springer, 2024](https://link.springer.com/chapter/10.1007/978-3-031-70368-3_21 )).
Building on prototype-based matrix factorization, the group has developed more explainable and bias-adjusted recommendation systems ([Best Student Paper, ECML-PKDD 2024](https://link.springer.com/chapter/10.1007/978-3-031-70341-6_4 )).
Deep and Explainable Rule Learning
The Computational Data Analytics group is known for foundational work in symbolic machine learning, particularly inductive rule learning. Current research develops algorithms that induce deeply structured logical IF-THEN rules capable of discovering intermediate concepts, investigated through approaches ranging from building structures from scratch ([Beck et al., DS 2024](https://doi.org/10.1007/978-3-031-78980-9_4 )) to principled optimization using SAT solvers ([Seip et al., SYNASC 2025](https://doi.org/10.1109/SYNASC69064.2025.00024)) and mixed integer linear programming.
The LORD rule learner delivers per-example explanations analogous to post-hoc XAI methods while scaling to datasets where established competitors fail ([Huynh et al., Machine Learning, 2023](https://doi.org/10.48550/arXiv.2301.09936 )).
Machine Learning in Game Playing
The Unit advances ML for strategic game playing, using contrastive learning to handle partial observability in games such as Reconnaissance Blind Chess ([IEEE Transactions on Games, 2024](https://doi.org/10.1109/TG.2024.3425803 )). Award-winning work on zero-shot card selection in collectible card games ([Best Paper, IEEE CoG 2024](https://doi.org/10.1109/CoG60054.2024.10645602 )) and ongoing efforts toward a foundational model for chess through self-supervised representation learning ([REPAIR, IEEE CoG 2026](https://arxiv.org/pdf/2606.11860 )) further demonstrate the group's contributions.
Ongoing Research Projects
FWF BILAI – Excellence Cluster (2024–2029): A nationwide Austrian effort led by JKU to integrate symbolic and subsymbolic AI into next-generation neuro-symbolic solutions ([bilateralai.net](https://www.bilateral-ai.net/home )).
ERC Advanced Grant "Whither Music?" (2022–2027): Investigating the future of music perception and understanding through AI ([Project page](https://www.jku.at/en/institute-of-computational-perception/research/projects/whither-music /)).
FWF doc.funds.connect "Human-Centered Artificial Intelligence" (2022–2027): A doctoral training program focused on trustworthy, human-centric AI ([dfc.hcai.at](https://dfc.hcai.at/)).
Collaborating across ELLIS
A flagship example of the Unit's cross-network engagement is its active involvement in the ELLIS Machine Learning for Molecule Discovery (ML4Molecules) program, one of the ELLIS network's key interdisciplinary initiatives at the intersection of machine learning, chemistry, and molecular sciences.
Established in 2023 and directed by Francesca Grisoni (Eindhoven University of Technology), Miguel Hernández-Lobato (University of Cambridge), and Nadine Schneider (Novartis), the program brings together Fellows and Scholars from ELLIS Units in Cambridge, Berlin, Linz, Lausanne, and Zurich, alongside leading researchers from industry partners including Novartis, AstraZeneca, and Microsoft Research.
The Unit participates actively in the ELLIS PhD Program, with doctoral students co-supervised by researchers at ELLIS Units including the University of Oxford (Yarin Gal) University College London (Marc Deisenroth) and the ELLIS Alicante Unit Foundation (Nuria Oliver), fostering cross-border mentorship and mobility within the European AI research landscape.
The ELLIS Unit Linz maintains active collaborations across the ELLIS network. Johannes Fürnkranz for example has a long-standing collaboration with Eyke Hüllermeier from ELLIS Unit Munich, with joint foundational contributions to preference learning ([Springer, 2010](https://doi.org/10.1007/978-3-642-14125-6)) and multi-label classification ([Machine Learning, 2008](https://doi.org/10.1007/s10994-008-5064-8); [ECML-PKDD, 2020](https://doi.org/10.48550/arXiv.2006.13346)), with current work focusing on uncertainty quantification for inductive rule learning.
Support for Young Talents
JKU Linz offers a comprehensive AI education pipeline, with dedicated Bachelor's, Master's, and PhD programs in Artificial Intelligence that attract students from all over the world. Enrollment numbers have been rising quickly, reflecting the growing international reputation of JKU and the ELLIS Unit Linz as a premier destination for AI education and research.
Supporting early-career researchers is a core commitment of ELLIS Unit Linz. The Unit currently hosts roughly 30 PhD students and postdoctoral researchers engaged in cutting-edge AI research. JKU has recently hired four female tenure-track postdocs in the areas of Neuro-Symbolic AI, Knowledge and Data Processing, Reinforcement Learning, and Natural Language Processing, strengthening both the Unit's research breadth and its commitment to diversity.
Through participation in the ELLIS PhD Program, the FWF doc.funds.connect program on Human-Centered AI, and the BILAI Excellence Cluster, the Unit provides structured training, international co-supervision, and extensive networking opportunities for the next generation of AI leaders.
Efforts in Public Engagement
ELLIS Unit Linz actively promotes dialogue between AI research and society through creative outreach and science communication:
Ars Electronica 2026: The interactive exhibit "Breaking News: Wolpertinger Sighted in Linz!" explored the mechanics of misinformation through an engaging public experience, receiving the Kepler Award in Science Communication 2026 (Video: https://www.youtube.com/watch?v=qgMErUgSxf0).
The Unit engages regularly with policymakers, industry, and the general public through open benchmarks, spin-off activities (NXAI, Emmi AI), and participation in Austria's national AI discourse.
View the highlights summary:
Text written by ELLIS Unit Linz Coordinator Yvonne Hauser.
