Interest in artificial intelligence is driving a proliferation of research into the nature of intelligence. Researchers at SFI are using it as an occasion to revisit classic problems and to make progress on some frontier questions around complex-adaptive systems.
“We have all these cases where some complex whole that displays intelligent capabilities is built up of parts that are themselves intelligent,” says SFI External Professor Jacob G. Foster, a professor of informatics and cognitive science at Indiana University Bloomington. “How do you build smart wholes from smart parts?”
The success or failure of collective intelligence can make or break a society’s response to challenges requiring large-scale coordination, such as climate change and COVID-19. A better understanding of intelligence could help institutions coordinate more effectively in the future, says SFI Professor Melanie Mitchell.
To explore these questions, Foster and Mitchell convened an SFI working group on March 19–20. This was the first in-person meeting of leaders on “Building Diverse Intelligences Through Compositionality and Mechanism Design,” a project funded by the Templeton World Charity Foundation.
The working group explored themes of compositionality — a property of parts that means they play well together — and mechanism design, or how parts can be made to interact and overcome incentive conflicts to produce a desirable collective behavior.
With expertise in fields like biology, computer science, engineering, and cognitive science, the researchers working on this project explore intelligence from different angles using a diverse array of formal models and model systems. They are united around central questions like, “What is the compositional structure — and what are the mechanisms — for taking these pieces and bringing them together in ways that make the whole more than the sum of the parts?” says Mitchell.
Compositionality is especially relevant in exploring AI systems’ ability to solve problems using abstraction and analogy. “AI has problems with generalizing outside of what it’s been trained on,” says Mitchell, whose group is exploring how AI can combine disparate pieces of knowledge in novel scenarios.
When it comes to agentic AI, having “a wider set of design principles for thinking about how to coordinate the behaviors of artificial agents, or humans working with artificial agents,” would be beneficial, Foster adds.
In addition to artificial intelligence, group members are investigating model systems such as scientific institutions, brains, evolutionarily designed soft robots, and communities that work together to solve puzzles.
The group plans to synthesize their knowledge in position papers that put forth formal frameworks for understanding compositionality and mechanism design, with the goal of offering researchers a lens they can apply to new systems of interest. They intend to reconvene next year to share progress.
Gathering to explore their research in-depth allowed attendees to identify overlaps in their work and stimulate collaboration. “Everyone I talked to came away from this meeting with a real buzz from the excitement of talking about these ideas together,” Foster says.