Scientists stress the need to grasp how artificial intelligence (AI) systems deal with unknowns if they are to fit into areas like medicine, transportation, and choices people make every day. The scientists argue that these systems must handle surprises well for smooth work alongside humans. A recent paper published in Nature Machine Intelligence shows that machines draw conclusions from what they know in ways unlike people. This gap matters a lot for future partnerships between humans and ai.
Generalization means taking lessons from familiar cases and using them wisely on fresh problems. For humans, it draws on ideas and pulling out main patterns from details. In AI, there are many generalization methods. One is out-of-domain generalization, where machine learning models trained on certain data try to work on totally new kinds. Another uses fixed rules in symbolic systems. Then there is neuro-symbolic AI, which mixes brain-like neural networks with strict logic to boost reasoning.
The scientists note that the term generalization shifts meaning between ai and human views. To fix this, they built a common structure. It looks at three main angles. First, what exactly counts as generalization. Second, the steps to make it happen. Third, ways to check if it works well.
Socio-technical systems
This work comes from a group effort by over twenty specialists from top schools in Bielefeld, Bamberg, Amsterdam, and London. It kicked off at a workshop held at the Leibniz Center for Informatics. The goal is to link cognitive science, which studies thought processes, with AI research. By spotting overlaps and splits, designers can craft systems that match human goals and ways of deciding better.
The study fits into the SAIL - Sustainable Life-cycle of Intelligent Socio-technical Systems project. Socio-technical systems blend people, tech, and society. SAIL focuses on making AI clear to see inside, and centered on humans from start to end.