Last ten frontier research projects launched thanks to AiNed call XS Europe
The latest 10 frontier research projects have been approved from the AiNed XS Europe call of the AI Coalition for the Netherlands (AIC4NL). These projects strengthen the AI knowledge and innovation base that is of great importance to the Netherlands. An important role is the connection of Dutch researchers worldwide, and especially in Europe.
The awarded research projects involve collaboration with several foreign collaborative partner organizations such as: Dresden University of Technology, Spanish National Research Council, Grenoble Institute of Technology, TU Wien, ETH Zurich, and Maxeler Inc. AI. Through these valuable collaborations, healthcare can lead to faster and better diagnoses. Another example is that through this collaboration, the quality of life can be improved for people who have poor communication with their loved ones due to paralysis.
The Netherlands faces the challenges of taking full advantage of AI opportunities. With AiNed XS Europe, space is created to give promising ideas, innovative and high-risk AI initiatives a chance.
Background AiNed XS Europe calls
This is the final round of grants awarded in 2024 within the AiNed call XS Europe. The call was intended for ideas and initiatives that focus on one or more challenges from sections 3 and 4 of the national AI research agenda AIREA-NL and are designed in cooperation with at least one foreign European cooperation partner organization. It is clear in advance whether the intended objective will be achieved. What matters is that every result, whether positive or negative, helps science move forward. The applications are anonymous, so the assessment is based purely on the research proposal and the European cooperation partner.
Want to know more about the projects?
We are happy to share more information about the assigned projects:
1. BCI-Found: BCI Foundational model for robust, generalizable and versatile neural implant performance.
Dr. Y. Berezutskaya, UMC Utrecht
People with severe paralysis cannot communicate with family and loved ones because of their inability to move and speak. Brain-computer interfaces (BCIs) can help by converting their brain activity into communication signals. However, current BCIs lack robustness, generalize poorly between users and support only limited communication strategies. BCI-Found offers a novel solution by using self-supervised training and transfer learning to create a fundamental model of brain signals. BCI-Found learns robust neural patterns that generalize across individuals and provides versatile communication options. If successful, it could transform BCI technology and significantly improve the lives of people with paralysis.
2. AI-SUSAT: Artificial Intelligence for Secure Underwater and SATellite Communications.
Dr. Y.C.G. Gültekin, TU Eindhoven
Reliable communications networks connecting national and regional entities are critical to the security of their citizens. An important aspect of these networks is communication with submarines. Satellite-to-submarine laser communication is therefore critical in the next generation of space networks. Due to the challenging nature of propagation through the atmosphere and salt water, traditional signal processing algorithms cannot support satellite-to-submarine laser communications. This project will develop data-driven algorithms to predict channel conditions and optimize laser signals. These algorithms will be the first to connect submarines to the mainland and complement global space communications networks.
3. SaTSNASP: Safe and Transparent Scheduling with Neuro-Symbolic Answer Set Programming.
Dr. J.L.A. Heyninck, Open University
Scheduling rosters is a pertinent problem. Needs change dynamically, and ideally rosters adapt dynamically. Current approaches for automatic roster generation are either logic-based, meaning they are verifiable but static, or they use machine learning, meaning they adapt dynamically but cannot be verified to meet all constraints. This project applies neuro-symbolic AI, a combination of logic-based and machine learning-based approaches, to develop AI tools that generate dynamic grids that are verifiable (in collaboration with a Dutch hospital and TUWien).
4. CONTACT-AI: CONTact in ACTion through Active Inference
Dr. W.M. Kouw, TU Eindhoven
Physical interaction with the real world is a huge challenge in artificial intelligence (AI). CONTACT AI explores new probability-theoretic techniques, based on computational neuroscience, to let walking robots explore by touch when vision provides unclear information. Explanatory models are used to allow an intelligent robot to switch between walking and dynamic interactions with obstacles. The intended implementation (Bayesian machine learning via information flows on factor graphs) is computationally efficient enough to run on small micro-computers aboard the robot. CONTACT AI provides methods for embodied KI systems that improve the market for walking robots in terms of robustness and autonomy.
5. ETAPE: Embodiment- and Task-Aware Parameter Embeddings for Robotic Foundation Models.
Dr. K.S. Luck, Vrije Universiteit Amsterdam
Much progress has been made in recent years in training large neural networks in the form of "robot foundation models" to "pre-train" and control different robotic arms and platforms; based on camera images and example data alone. However, direct motor control, such as angular positions and speeds, remains a challenge for these models when deployed on different robotic platforms. This proposal explores new ways for modularizing and structuring the so-called "embedding space" of robot foundation models, allowing them to better identify and generalize the design and task of the robot.
6. AI-Driven Cancer Diagnostics: Explainable and Transparent AI Tools for Personalized Treatment.
Dr. M. Menzel, TU Delft
To improve the survival rates of cancer patients, personalized treatment immediately after surgery is crucial. However, histological tissue analysis can take several weeks. This project aims to create explainable and transparent AI tools for cancer diagnostics. Using a large dataset of labeled tissue samples, an AI will be trained to distinguish tumor from healthy tissue and to assess cancer severity during surgery by analyzing collagen fiber organization at tumor boundaries. If successful, the AI tools developed will enable earlier diagnosis and improve survival rates for cancer patients, with users able to understand how the algorithm arrived at its decision.
7. QP-GPT: A foundation model for quantitative perfusion MRI.
Dr. C.M. Scannell, TU Eindhoven
For many diagnoses, the amount of blood flow, or perfusion, reaching the organ is crucial to clinical decision making. A promising idea is to use AI to model perfusion in MRI scans instead of relying on the physician's subjective interpretation. This project improves AI model training for quantifying perfusion by creating a foundation model that can be quickly adapted to specific patient data. This makes better use of large data sets without expert annotations and integrates physical laws, resulting in faster and more reliable decisions for patients.
8. MBFM: A Multimodal Brain-Signal Foundation Model for unified brain analysis and disease diagnosis.
Dr. C. Strydis, Erasmus MC
This project aims to develop a multi-modal brain signal foundation model (MBFM) that can integrate different brain data types such as brain activity, EEG and video images. Foundation models are typified by their general and broad applicability for various subtasks and rapid deployment for specific subtasks, i.e. diagnostics of neurodegenerative diseases, decoding brain signals in experimental research or developing brain-computer interfaces. The ambition is to achieve this by extending existing foundation-model architectures for brain signals. Although the ultimate focus is on humans, the MBFM is being trained as a proof-of-concept on publicly available rodent datasets and internal clinical data, supported by Groq's AI infrastructure for efficient processing.
9. Brain-inspired MatMul-free Deep Learning for Sustainable AI on Neuromorphic Processor
Dr. G. Tang, Maastricht University
Deep learning relies on energy-intensive matrix multiplication computations (MatMul) on GPUs, which become unsustainable as neural networks grow larger. Inspired by the brain's efficient, asynchronous and local computations, this project aims to develop a new computational paradigm for deep learning that reduces energy consumption and latency, and moves away from traditional GPU-based MatMul. In collaboration with researchers at TU Dresden, this project aims to implement brain-inspired computational paradigm on neuromorphic processors and integrate it into general purpose tools. This approach has the potential to make AI more sustainable and accessible to large-scale applications, reducing energy costs and environmental impacts.
10. Multimodal Representation Learning for Evolving Cardiac State and Risk Estimation.
Dr. F.V.Y. Tjong, Amsterdam UMC
By combining different types of medical data, medical AI systems can track a patient's health over time, determine the need for specialist checkups and predict future problems. However, current AI tools struggle to process incomplete or inconsistent medical data and cannot accurately predict how a patient's health status changes over time. Efforts are underway to create a multimodal AI system that can combine different types of diagnostic information (cardiac ultrasounds, MRI images, patient data) to provide a more complete picture of a patient's health status and thereby detect problems early.
