Dice Question Streamline Icon: https://streamlinehq.com

Ascertain the implications of learning from naturalistic data in AI for natural intelligence

Ascertain the implications for natural intelligence of the observed positive interaction between learning-based systems and naturalistic data in modern AI, wherein models trained on naturalistic data often show qualitatively different generalization than those trained in simplified settings.

Information Square Streamline Icon: https://streamlinehq.com

Background

The authors highlight a recurring empirical trend in AI: learning-based systems not only accommodate naturalistic data but also benefit from it, exhibiting qualitatively different generalization behaviors compared to training in simplified environments.

They explicitly note that the implications of this trend for natural (biological) intelligence are currently unknown, framing a key open direction for cognitive science to investigate.

References

One common trend is that there is a positive interaction between learning-based systems and naturalistic data—learning-based systems can accommodate naturalistic data, and reciprocally, learning from naturalistic data results in qualitatively better results than learning in simpler settings. While we do not yet know the implications for natural intelligence, they open interesting questions and challenge prior assumptions.

— Naturalistic Computational Cognitive Science: Towards generalizable models and theories that capture the full range of natural behavior (2502.20349 - Carvalho et al., 27 Feb 2025) in Section 4: The (surprising) benefits of learning with naturalistic data