Almost halfway the NWO Gravitation programme Hybrid Intelligence

The Gravitation programme Hybrid Intelligence has been running for nearly four years now, and there are still six more years to go. In the latest issue of the I/O magazine from IPN (ICT Research Platform Netherlands) Hayley Hung and Bart Verheij (Multi-Agent Systems) were interviewed to evaluate what has been accomplished so far, wat will be the implications of ChatGPT and how they hope the programme will affect the development of AI.

The first batch of PhD students has nearly completed their research, and recruitment of the second batch has begun earlier this year. This also means that the first concrete research results will soon be published. Hayley Hung and Bart Verheij can already share some preliminary highlights.

One concrete research question that Bart has explored with PhD student Cor Steging is how to make the outcomes of machine-learning models more transparent and explainable. Bart states: ‘We have looked at a rule of Dutch law that defines exactly what and when something qualifies as a so-called tort, a legal topic about compensation for damages. With a knowledge representation model of that law, we have generated data that we have fed into a machine-learning model. We asked the question whether a machine-learning model could rediscover the underlying knowledge structure from the data.’

It turned out that the machine-learning model performed very well statistically, in the sense of mostly giving the right answers, but not for the right reasons. Verheij: ‘It failed to discover the underlying knowledge structure. This, of course, is very strange and undesirable. The next challenge is to find out why that is and what we can do about it. For now, machines cannot do the job by themselves, and having humans in the loop is necessary.’

Please read more in the I/O magazine (pages 5-7) about for instance how Bart considers large language models to be a game changer.