‘The future is interdisciplinary’

Kerstin Bunte, Professor of Machine Learning for interdisciplinary data, combines data with expert knowledge to solve complex problems. This is particularly valuable in fields where data are scarce, such as paediatrics. In this episode of the interview series JTS Scholars, Bunte talks about her work, in which she brings together different disciplines. ‘I want to be an ambassador for true interdisciplinarity.’

Text: Jelle Posthuma

Machine learning is a field of AI in which computers learn patterns from data in order to make predictions or decisions. Many well-known AI tools are trained on enormous amounts of data. However, in some situations, only limited data is available. It is precisely this problem that Professor of Machine Learning Kerstin Bunte focuses on: how to build reliable models even with limited data. An important principle for her is the use of expert knowledge. ‘If you have a complex problem, you can collect more data, but that is not always feasible. And sometimes the necessary information simply cannot be obtained. The use of expert knowledge provides a good alternative. This is also where universities can make a difference, because we do not compete with companies like Google and Amazon.’

From trial and error to deeper understanding

One important application area is (bio)medical science, where doctors possess extensive expert knowledge, but it is virtually impossible to collect large amounts of training data from a single patient. Therefore, researchers often work with data from patient populations. ‘Perhaps 80 to 90 percent of the population falls within that group, and for them medication works well. But for the remaining 10 to 20 percent, it does not work, or works less effectively.’

In paediatrics in particular, the lack of data is a major problem. ‘Children are not small adults: more changes occur in the first 15 to 16 years of life than in the entire period from adolescence to the end of life.’ Due to this rapid development, children relatively often receive under- or overdoses of medication. There is simply very little data available on children across different age groups, especially compared to adults. And this is especially true for healthy children. ‘That is also understandable: we cannot expect millions of children to provide data through invasive methods.’

The lack of data leads to trial and error in medicine, as Bunte knows. ‘I experienced it myself a few years ago, when I was being treated with blood thinners. For two weeks, I had to have blood taken every morning before work to check whether they’d got the dose right. I thought: this is stupid. Surely there must be an easier way?’ For Bunte, this is an important motivation to collaborate with medical experts on more personalised methods. ‘Doctors are often highly motivated to improve things, especially in paediatrics. By combining knowledge of biological processes with machine learning, we can better understand why medicines do not work for some groups.’

A wide range of applications

Bunte’s collaborations are not limited to medical science. There are many other application areas, the professor explains. She mentions the smart industry, where production processes have become so small and compact that it is not always possible to add sensors; here too, available data are very limited. ‘But we do know what goes in and what comes out, and based on expert knowledge you can form a good understanding of what is happening inside the machine. We can build a model of that, which is crucial when something goes wrong, because it allows us to trace the cause.’

Machine learning is also useful in astronomy. In this field, there is no shortage of data, but much of it is unusable. ‘Ninety-nine percent of the universe is empty. For everything that is bright and visible, there is sufficient data, but other phenomena remain more hidden. We often see snapshots in time: very long processes of which we only observe a single moment.’ Yet, with knowledge of the physics we can formulate plausible explanations, Bunte states. ‘Based on that, we create simulations, which are then run using algorithms.’

Interdisciplinarity requires investment

According to Bunte, combining different disciplines is challenging. ‘You first have to learn to speak the same ‘language’ and immerse yourself in the other field. When a machine learning researcher talks about a model, they typically mean a data-driven prediction model; an engineer might refer to a simulation or mechanistic model; and a medical doctor could be talking about an organism, such as a mouse model. And don’t even get me started on the abbreviations! That makes writing papers and proposals even more complicated,’ she laughs. The way of thinking also differs. ‘But that is exactly what makes it fascinating. You miss out on so much knowledge if you only work monodisciplinarily: the future is interdisciplinary.’

Bunte aims to contribute to this as a JTS Scholar. ‘I want to be an ambassador for true interdisciplinarity. Right now, it is sometimes still artificial. Interdisciplinarity must have added value, and that requires investment. In my view, research is truly interdisciplinary when both fields publish about it. Take our collaboration with astronomers, for example. Both the machine learning community and the astronomy community published papers on this. That is a good indicator. It is complementary: machine learning is not just a smart tool, but together with expert knowledge it forms its own contribution.’