summary: Humans are excellent at adapting to new situations, but machines often stumble. New interdisciplinary research reveals that the roots lie in the process of how humans and AI approach “generalization,” that is, transferring knowledge to new problems.
Humans rely on abstraction and conceptual frameworks, but AI systems apply statistical or rule-based methods, each with limitations. Bridging these approaches open the way for a more flexible, human-centric AI system that better adapts to the complexities of everyday life.
Important facts
- Different meanings: “Generalization” brings a variety of definitions in the research of cognitive science and AI.
- Human vs AI: Humans generalize through abstraction. AI uses domain-specific processes.
- Shared Framework: Researchers propose a unified framework to better coordinate human and machine reasoning.
sauce: Bielefeld University
How can humans adapt completely to new situations, and why do machines struggle with this so often?
This central question is explored by researchers in Cognitive Science and Artificial Intelligence (AI) in a joint article published in the journal.Nature Machine Intelligence”
“If you want to integrate AI systems into your daily life in medicine, transportation, and decision-making, you need to understand how these systems handle unknowns.”
“Our research shows that machines generalize in a different way than humans, which is essential for the success of human-AI collaboration in the future.”
The difference between humans and machines
The technical term “generalization” refers to meaningful conclusions about unknown situations from known information, the ability to flexibly apply knowledge to new problems.
In cognitive science, this often involves conceptual thinking and abstraction. However, in AI research, generalization serves as an umbrella term for a variety of processes, from machine learning beyond known data domains (“outside domain generalization”) to rule-based inference in symbolic systems, to so-called neural symbolic AI combining logic and neural networks.
“The biggest challenge is that 'generalization' means that it's completely different for AI and humans,” explains Benjamin Paaßen, a junior professor of knowledge representation and machine learning at Bielefeld.
“That's why it was important to develop a shared framework. What does generalization mean in line with three dimensions? How is it achieved? And how can it be evaluated?”
The importance of AI for the future
This publication is the result of the interdisciplinary collaboration of more than 20 experts from leading international research institutes, including the University of Bielefeld, Bamberg, Amsterdam and Oxford. The project began with a joint workshop at Schloss-Dagstour's Leibniz Informatics Centre, co-organized by Barbara Hammer.
The project also highlights the importance of bridging cognitive science and AI research. Only by a deeper understanding of these differences and commonalities can we design AI systems that can better reflect and support human values and decision-making logic.
This research was conducted in the sustainable lifecycle of the collaborative project Sail-Intelligent Socio-Technical System. Sail explores how AI is designed to be sustainable, transparent and human-centric throughout its lifecycle.
Funding: The project is funded by the Ministry of Culture and Science, North Rhine Westphalia.
About this AI research news
author: Jorg Healen
sauce: Bielefeld University
contact: Jörg Heeren – Bielefeld University
image: This image is credited to Neuroscience News
Original research: Closed access.
“Adjusting generalizations between humans and machines” by Barbara Hammer et al. Nature Machine Intelligence
Abstract
Adjusts generalization between humans and machines
Recent advances in artificial intelligence (AI), including generative approaches, have brought technologies that can support humans in the formation of scientific discoveries and decisions, but can disrupt democracy and target individuals.
Responsible use of AI and participation in human and AI teams increasingly demonstrate the integrity of AI, the need to make AI systems according to their preferences.
An important but often overlooked aspect of these interactions is the various ways humans and machines generalize. In cognitive science, human generalization generally involves abstraction and conceptual learning.
In contrast, AI generalizations include out-of-domain generalizations in machine learning, rule-based inferences for symbolic AI, and neural-to-membrane AI abstraction.
Here we combine AI and cognitive science insights to identify important commonalities and differences across three dimensions: concepts, methods, and evaluation of generalization.
Along these three dimensions, we map different conceptualizations of generalization in AI and cognitive science, taking into account the role of alignment in human and AI teams.
This creates interdisciplinary challenges in AI and cognitive science, and needs to be addressed to support effective and cognitively supported alignments in human-AI teaming scenarios.
