We constantly hear about the -AI revolutionand the -digital transformation-. This is being driven by rapid advances from the field of machine learning (ML), such as those enabling autonomous vehicles, ChatGPT, DeepSeek and even predicting extreme weather events. But how do these innovations move from fundamental research to real-world impact? Geert-Jan Houben explores this question with the co-directors of the ELLIS Unit Delft: Frans Oliehoek and Jens Kober.
How -machine learningfits within -artificial intelligence (AI)-
-AI has a tradition of formulating problems as search or optimisation problems. However, this means we know what we are trying to optimise, and that we have the right data to estimate probabilities. But, in many circumstances we don’t know what we are trying to optimise or we don’t have this data,- explains Frans Oliehoek, Associate Professor at TU Delft’s Faculty of Electrical Engineering, Mathematics and Computer Science (EEMCS).
-That is when machine learning comes into play as a subfield of AI,- continues Frans. -ML uses algorithms to learn from data, which assures that our solutions are a better fit for real world problems. For instance, ML can be used to predict rare events like extreme weather. By finding patterns in large or complex datasets, ML models offer new insights where traditional methods may fall short.-
-Besides, ML is now an integral and fundamental part of many AI systems. In autonomous driving, for example, rules-based models fail in complex and unpredictable environments like crowded urban areas. But ML can learn from observations, adapting to unpredictable scenarios like these.- Jens Kober adds, Associate Professor at TU Delft-s Faculty of Mechanical Engineering (ME). -An important benefit of ML is that it can automate tasks, such as face recognition, that previously were done manually because we don’t understand them well enough to program solutions explicitly.-
In brief, machine learning focuses on finding patterns in complex data, and readily adapting to new or unpredictable environments - which combined can offer novel insights for a variety of applications.
Geert-Jan Houben
Using machine learning in the real world
-Our fundamental machine learning research is a key enabler of research and innovation. The insights of such -in-AI- or -in-ML- research is increasingly embedded in our -with-ML- research across campus, driving novel approaches in the lab and advancing impact in domains ranging from individual healthcare to mobility and energy infrastructures. In this way, we do ML for the real world,- highlights Geert-Jan Houben, Pro Vice Rector Magnificus AI, Data and Digitalisation and leader of the TU Delft AI Initiative.
-A driver of our -in-ML- machine learning is the ELLIS Unit Delft, which Jens and I co-direct. Members of ELLIS are internationally recognised experts in ML, and they are working on a very broad spectrum of ML techniques and application domains. At the ELLIS Unit Delft, we have a strong focus on connecting theory to applications. I think that this sets us apart compared to some other ELLIS units,- explains Frans. -In addition to that, applying ML to particular problems isn’t always straightforward. In urban planning, for example, ML can be used for predictive maintenance of critical infrastructures, but the few available public datasets do not capture the richness of the problem.
-In healthcare, as another example, ML can help with detecting abnormalities in patients that may go otherwise unnoticed for a variety of reasons. These uses bring benefits, but we also need to ensure the underlying systems are reliable and understandable. So, it’s crucial that engineers tackling problems -with ML- collaborate in developing and applying fundamental methods,- adds Jens.
Frans continues: -What these examples show is that being an integral part of the broader AI ecosystem helps our ELLIS Unit to spread novel ML insights into engineering applications, while at the same time feeding back the limitations of existing methods into the ML community.-
Our ambition with the ELLIS Unit Delft is not just to achieve academic success in ML, but also to translate theoretical research and fundamental techniques into practical solutions that address urgent societal challenges.
Frans Oliekoek en Jens Kober
Geert-Jan: -I would add that an overarching theme of these examples is that they show how interdisciplinary research and innovations can be done right, not just fast. By -rightwe mean it’s good science, but also aims to do some good. For us, this is at the core of ML for the real world.-
Translating machine learning into the real world
-Translating ML to real-world applications is far from trivial though. Models will never be perfect, but with the data we have, we can begin to bridge the gap by testing how techniques translate across different domains and situations. Discussing what makes a good model helps scientists not only solve problems but also better understand what-s needed.- says Jens.
-For example, a company we collaborate with developed a neural network to predict salt intrusion in canals. While it performs well on their dataset, questions remain about its robustness to changes like climate shifts or new infrastructure. The more a model can adapt to real-world changes the better, but achieving this is challenging,- Frans explains. -This is where iterative research and design become crucial; Models aren-t perfect from the start; they improve through refinement and testing in real-world conditions. Collaboration with scientists in other fields provides the necessary context to test approaches, since the real world doesn’t behave like a controlled lab. For instance, while a computer vision model might identify stop signs well in clean datasets, stop signs in the real-world are often damaged, dirty or obscured, which poses a challenge that strengthens model robustness. Our ELLIS researchers bring significant expertise on such issues of robustness, as well as iterative ways of dealing with this.-
Geert-Jan highlights that another essential consideration is understanding the boundaries of ML. -Indeed, not every problem is suitable for these methods,- agrees Jens. -Sometimes, the available data doesn’t meet the requirements for ML to work effectively, or the objective of the problem is unclear, which can lead to unforeseen issues like discrimination or failure to take into account human preferences. In such cases, advising against the use of ML is just as important as applying it, ensuring that the technology is used responsibly and effectively.-
Finding the right balance between regulation and innovation
-Sometimes, optimisation goals in machine learning don’t align with societal benefits. For instance, recommender systems often prioritise engagement, but this can lead to filter bubbles that limit diverse perspectives,- mentions Jens.
-I think this highlights the need to think critically about what we optimise and why. Moreover, human decision-makers remain essential. ML cannot fully replace human judgment, particularly in complex systems that demand meaningful human control,- highlights Geert-Jan.
-This consideration extends to regulation,- adds Frans. -Developing responsible AI requires an iterative process that allows for real-world testing while addressing ethical considerations and implementing safeguards. Scientists play a critical role in ensuring the proper use of AI, particularly in sensitive areas like healthcare, autonomous driving, and traffic management. By overseeing both the technology and its responsible application, they help balance the competing demands of innovation and regulation.-
Achieving balance between regulation and innovation is vital for responsible AI development.
Frans Oliehoek
-Yes, while space for experimentation is necessary to drive innovation, it must be accompanied by robust mechanisms to ensure safety and uphold ethical standards,- stresses Geert-Jan.
The power of interdisciplinary collaborations in ML
-ML is becoming pervasive and ubiquitous in engineering and design, as well as an essential tool in scientific labs. It means we see that ML increasingly acts as the -enginedriving societal impact as well as scientific research. This makes collaboration and integration essential,- explains Geert-Jan.
-The image of a lone scientist striving to achieve a major break-through on their own is inherently outdated,- Jens agrees. -Research heavily relies on collaborations and (international) networks. Having a strong local network allows us to offer -the complete packageincluding both algorithmic and domain expertise, while being part of the ELLIS network allows us to collaborate with experts from across Europe.-
Collaboration is essential for testing and refining approaches in real-world scenarios, ensuring they deliver meaningful results.
Jens Kober
-ML enhances the tools available to other disciplines, fostering collaboration. Researchers gain from connecting with experts to effectively apply these techniques in practice. Collaboration with domain experts helps us understand what-s needed and refine our methods. That-s how we advance both science and its practical impact,- adds Frans.
ELLIS - the European Laboratory for Learning and Intelligent Systems - is a pan-European AI network of excellence focusing on fundamental science, technical innovation and societal impact. Founded in 2018, it now includes 43 sites across 17 countries and over 1.500 active researchers. In addition to building a network of top machine learning researchers across Europe, ELLIS has a mission of advancing ethical and trustworthy AI for positive social impact.
The ELLIS Unit Delft has almost 40 members from across four faculties. Check the website or the weekly AI Update for more information about the Unit-s research and upcoming events.