Why concepts are (probably) vectors

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Debate on the nature of human concept representation.
• Requirement for representations to support diverse cognitive computations (similarity, categorization, relations).
• Need for representations to enable theory development and procedural knowledge.

Purpose of the Study:

• To argue for vector-based representations as a unified account of human concepts.
• To highlight the compatibility of vector representations with neural architectures.
• To discuss recent advances supporting this view.

Main Methods:

• Conceptual analysis of representation requirements in cognitive science.
• Review of recent advancements in large language models (LLMs).
• Examination of vector symbolic architectures (VSAs).

Main Results:

• Vector representations can account for a wide range of cognitive properties (similarity, features, categories, definitions, relations).
• Vector representations support complex cognitive processes like theory development and ad hoc categorization.
• Recent LLMs and VSAs demonstrate practical implementation of vector-based symbolic computation.

Conclusions:

• Vector-based representations provide a compelling and neurally plausible model for human concepts.
• Emerging AI technologies validate the power of vectors for symbolic processing and cognitive modeling.
• This approach unifies diverse cognitive functions under a single representational framework.