‘We were seen as programmers rather than proper researchers’
Twenty years of AI research at the Faculty of Science
As a young researcher in Utrecht, he focused on studying artificial intelligence when this was little more than abstract science. Now he is a professor and AI has found its way into the very core of society as well as the university. Twenty years of AI research at the Faculty of Science, seen through Mehdi Dastani's eyes.
When Mehdi Dastani started in Utrecht in 2001, Utrecht University boasted a first in the Netherlands: it was the first Dutch university to set up an undergraduate programme in cognitive artificial intelligence. However, Dastani, a postdoc newly arrived from Amsterdam, would not be teaching there: the course came under Humanities, “where people were interested in the philosophical issues surrounding artificial intelligence”, whereas Dastani became a researcher with the Faculty of Science.
Dastani, too, had a background in philosophy, having studied it alongside his degree in computer science in Amsterdam. “I chose the artificial intelligence programme in my computer science degree, which soon led to Big Questions such as: what even is intelligence? Are computers actually capable of thought? Consciousness? I not only wanted to understand the technical aspects of artificial intelligence, but to grasp the underlying concepts and theoretical context.
Uncharted territory
In Amsterdam in the early 1990s, he was one of the few to delve into neural networks. ‘I was something of a maverick really, nobody believed in it. There was no data, no strong computers to work with. You had connectivity between a few dozen neurons at best. Essentially, it was unworkable.’ For this reason Dastani moved towards logical methods. In Utrecht, under the guidance of John-Jules Meyer, a pioneer in the field of AI in the Netherlands, he worked on multi-agent systems, a previously relatively unexplored area of computer science.
At Meyer's Intelligent Systems group, Dastani developed computer systems and software to communicate and interact to achieve certain goals. ‘Think of railways, for example: a single train can be controlled by a single system, but to share the rails and deal with delays, for instance, the systems have to communicate with each other and coordinate how to proceed. The same principle applies to self-driving cars.’
Dastani: ‘John-Jules Meyer was interested in AI from a perspective of logic: how can you build an intelligent system from logical models? Computer science has a strong basis in logic, because in effect a computer is a logical machine: every calculation can ultimately be traced back to a series of logical operations using ones and zeros. So being able to build more complex programmes using richer logic wasn’t really far-fetched. We thought: can we find a way where all you have to do is say, you have this knowledge and these goals, find a way to get to those goals?”
There was a limit to this: logical models cannot be endlessly extended with new rules that in turn create exceptions to other rules. As a result, the group experimented with specific machine learning techniques, such as reinforcement learning, an approach that was not widespread at the time. Here, the computer system does not rely on logic and formal rules, but rather learns from experience.
Artificial intelligence as a learning system wasn’t feasible due to insufficient data and computing power
The field of multi-agent systems was steadily gaining momentum; symposiums, conferences, new scientific journals. There was ever more interest. Dastani: ‘But in those early years, it was mostly an abstract notion. Artificial intelligence as a reasoning and learning system was ultimately unworkable because in reality there was too little data and too little computing power. So we couldn't show what was actually possible. At that time, we were not seen as a fully-fledged scientific field that also did fundamental research. The university saw us mainly as service providers, able to help other disciplines with programming.’
This situation changed around 2010, when increasingly powerful computational programmes became available, more data was accessible as well as more theoretical knowledge about neural networks. ‘Digitalisation had unlocked a huge amount of data; across all sciences, people were now working with large data sets. It took AI to analyse and use that data. Rather than having to work everything out in our own heads, we could instead use the data to train the systems to learn patterns and structures on their own.’ Artificial intelligence became intertwined with all disciplines of the university.
Dastani's reputation among colleagues as an omnivore in the field of AI has served him well. ‘’Throughout the years, I have delved into psychology, sociology, political science and economics. I look at the world and I want to understand what is happening. And then you see that AI touches on all those fields: if you look at theories of economics, for instance, that brings you to social choice theory and game theory. These are mathematical models which explain how choices are made. Political and social theories also lead to concepts like fairness and bias, which again are major themes in AI.”
He seeks to encourage this same curiosity towards the world around him within his group. “A scientist should be playful. You cannot bind yourself to a certain area or ideology, as that will limit you. Instead, sample everything. It is important to me that my researchers keep asking themselves: how can I use my experience for some new development? And then I hope they take their creative thoughts seriously. That's how I use my own baggage, too.”
Great benefits, great dangers
His background ensured that Dastani has always kept a broad perspective on AI. As a high school student in the 1980s, he abruptly left his homeland Iran where a war was raging with neighbouring Iraq. Eventually, he managed to reach Europe. “I experienced things that shaped my view of the world. It makes me alert to what is happening in the world at the moment. Digitisation and the advent of AI is a big social development; just consider the influence of big tech, the addictive nature of social media, the spread of disinformation and misinformation. AI can bring great benefits, but we observe that it also brings great dangers.”
Here, Dastani sees an important role for the university. We have such broad expertise, not only within the Faculty of Science but at other faculties as well. We need to tap into that if we want to remain leading in AI research. I actively advocate pooling all knowledge within the university in the field of AI as a powerful and coherent response to the challenges of our time.’
I have no regrets. But sometimes I dream about what could have happened if I’d taken neural networks seriously.
Dastani believes that currently this is happening insufficiently. Indeed, our knowledge about AI is even in danger of fragmenting across different faculties and departments, he observes. “I have been fighting for some time for somewhere to connect and support all AI researchers at the university.” Realising such a centre of expertise is not straightforward, notes Dastani. “I know when to be diplomatic, when it is important to weigh everyone's interests and needs. But sometimes you have to stand for something and convince others to go along, or you end up losing relevance.”
A professor at Intelligent Systems since 2018, Dastani succeeded John-Jules Meyer as chair of the group in 2022. His colleagues praise him for his drive, his courage and his need to understand everything in detail. In the field of multi-agent systems, he is among the world's best. “I have gained some great knowledge over the past few years and I have no regrets. But sometimes I think back to my master's thesis on neural networks. If I had taken it seriously then, when no one else did, who knows what that could have led to. I do dream about that sometimes.”
Five facts about Mehdi Dastani
- For many years, Mehdi Dastani has been collaborating with Jan Broersen, from Humanities. Together, they lead the focus area Human Centered AI
- Mehdi Dastani is a EurAI Fellow, a title awarded to a select group of individuals who have made significant and long-term contributions to the field of artificial intelligence in Europe
- Most universities have only one researcher in the field of normative multi-agent systems. Thanks to Mehdi Dastani, there are about five working in Utrecht who rank among the world's finest
- Mehdi Dastani is always up for a game of table tennis
- Mehdi Dastani is a great cook and he is happiest spending hours in the kitchen.
Mehdi Dastani chairs the Intelligent Systems research group, which holds considerable expertise in multi-agent systems. But that is not the only place at the Faculty of Science with a wealth of knowledge on AI and Data Science:
The Faculty of Science celebrates its 20th anniversary!
This article is created as part of the 20th anniversary of the Faculty of Science of Utrecht University. It is part of a series of (visual) stories highlighting this jubilee.