Can we trust machine learning? Henrik Boström’s advice to researchers and companies

Can we trust machine learning? Or AI in general? We asked Henrik Boström, one of Sweden’s leading researchers in machine learning.

Henrik Boström is a professor of computer science with a focus on data analysis systems at the Department of Software Technology and Computer Systems at KTH. His research is about machine learning algorithms, with a particular focus on confidence-based prediction, ensemble models, and explainable machine learning. He has led and worked on projects applying machine learning in the pharmaceutical industry, healthcare, automotive industry, and insurance sector. He has been an editor for the journals Machine Learning and Data Mining and Knowledge Discovery, and has served as a program committee member for some of the top conferences in the field for over two decades. Henrik Boström is also part of the committee of external experts providing feedback on the activities of AFAIR. This spring, he was one of the speakers at our event where he shared current insights on machine learning and reliability.

Hello Henrik! Can we get a quick introduction?

”I am a professor of computer science focusing on data analysis systems. My research focuses on reliable machine learning, especially explainability and prediction with confidence. I love programming and experimenting, and think it’s fantastic to be paid to pursue a hobby I have had since I was very young.”

Can we trust machine learning in particular and AI in general?

”The short answer is no, we cannot.”

Why not?

”There are many pitfalls in developing machine learning models that often make us overestimate their accuracy. There are countless examples of projects resulting in useless or sometimes directly harmful models, often due to a combination of poor methodology and an inability to understand the limitations of the models.”

What can we do about it?

”We can, for example, take courses that not only focus on the technical aspects but also thoroughly cover the methodological aspects.”

Your best advice to other researchers and companies?

”To researchers who have not yet collaborated with companies and organisations, I want to highlight the potential in such collaborations, not least to identify new and relevant scientific questions. To companies starting new AI projects, I would like to emphasise the value of collaborating with well-educated actors. It is not enough that they have the technical competence to build models and implement systems; they must also be methodologically competent to avoid the many possible pitfalls.”