Logic and machine learning together offer the best of both worlds

On Monday 30 May, Mehdi Dastani, professor of Intelligent Systems at Utrecht University, will deliver his inaugural lecture. In his research, he is working on a broad AI vision in which the focus is on how humans can remain in control of ever smarter systems. "The choices made by AI systems should be understandable, explainable and traceable for humans, if we want to keep control over these systems and trust them. Then AI systems will be at the service of humans instead of the other way around."

Artificial Intelligence is involved in many social and economic choices, including those with potential risks: in the medical world, AI systems make diagnoses based on digital image recognition, and they play a major role in the energy transition, mobility and financial markets.

Until now, such high-risk processes have always involved humans to monitor the choices made by the systems. "Human specialists are still needed to watch out for exceptions in order to avoid errors and risks," says AI researcher Mehdi Dastani. "Moreover, there are always humans who need to be accountable if things go wrong because an AI system makes a mistake."

As long as it is unclear how AI systems that are based on black-box algorithms work, society can't really have rock-solid confidence in them.

Black box algorithms

In order for smart computer systems to make increasingly risky choices in the future, without too much intervention from human specialists, we need to be confident that humans can maintain control over these types of systems, says Dastani.

"Recent AI developments - successes and failures - have led many to hope for or fear the future impact of AI technology. This is partly due to the fact that AI technology is increasingly based on learning from data using black-box algorithms. They do work - sometimes very effectively - and that is of course positive. But there are also risks attached to their use as it may not be clear which data properties the system uses to arrive at its decision. For example, if one person can get a lower mortgage to buy a house than someone else, you want to be able to determine the reason. But AI systems that are based on black-box algorithms do not always succeed in doing that, and as long as that is unclear, society can't really have rock-solid confidence in these systems."

Human-centered Artificial Intelligence

The goal of Dastani's research is to align the design of AI systems with human cognition and capabilities. "In a sense, AI systems should perceive, learn, reason, and act on the world as humans, for example, by ensuring that the features and rules they use to classify images match those of humans, and by allowing AI systems to reason about what they learn. If a system learns that portraits 1 and 2 show the same person, and portraits 2 and 3 do too, but portraits 1 and 3 do not show the same person, then something went wrong when learning to classify the portraits. By combining good old-fashioned AI with modern data-driven AI techniques, powerful human-centered AI systems can be built that we can understand and explain, by tracing and analysing the underlying reasoning and route of those decisions. These ingredients are needed to increase trust in AI systems."

The broad Utrecht University vision of AI is closely aligned with the European policy to stimulate human-centered Artificial Intelligence in Europe.

From small to big data

Prof. Mehdi Dastani has been with Utrecht University since 2001. "UU was the first university in the Netherlands to offer the Artificial Intelligence programme. At that time, the field of artificial intelligence, and therefore our programme, was driven by logic-based methods and insights from various scientific disciplines, which by the way were obtained by small data. In the meantime, artificial intelligence has developed enormously, mainly due to developments in machine learning, the availability of big data and the powerful computing capabilities. Our bachelor's and master's programmes in artificial intelligence have kept up with these developments and are now among the best in the Netherlands."

Together with Prof. Jan Broersen, Dastani set up the focus area Human-centered Artificial Intelligence at Utrecht University. Dastani: "With the Utrecht focus areas Governing the Digital Society and Applied Data Science, which deal with the social, ethical and legal aspects of digitalization and data, we have ensured that AI research in Utrecht has become human-centered. Under this umbrella, we stimulate UU-wide AI research and cooperate with national and international partners. This broad vision of AI is closely aligned with the European policy to stimulate human-centered Artificial Intelligence in Europe."