Elizabeth Spelke identifies six systems of core knowledge in humans and animals: objects, agents, persons, places, forms, and number. Each system is domain- and task-specific, operating as a distinct “unitary whole”; they are engaged under different conditions and capture different features of the world.
Core knowledge systems share the following properties:
- Centered on abstract, interconnected concepts;
- Limited in specific details about the world;
- Activate particular regions of the brain across different animals—core knowledge systems are ancient;
- Innate, present and functional at first encounter;
- Emerge early and are present throughout life—same signature patterns show up in adults.
According to Spelke, the ancient property of core knowledge gives rise to the other shared properties. For example, for systems to have persisted to the present day, they could not have centered on details of the environment—they were necessarily abstract and limited. The systems had to be localized in the brain, rather than distributed, to survive other evolutionary changes in the central nervous systems. Finally, given that ancient systems center abstract learning, they must be innate—otherwise, they would require immense amounts of data to extract the subtle information that captures abstract categories.
Question #open-question What implications does the evolutionary origin of properties imply for AI systems, particularly the final point about data? Is engineering this knowledge in silico too difficult? Can evolutionary algorithms be productive (see application-2024-sfi-ucr)?
Related: Evolution is the only guaranteed method of producing general intelligence