Upending assumptions about learning, inspired by an AI phenomenon (www.santafe.edu)

A recent study has challenged long-held beliefs about learning in both AI and humans by introducing the concept of “excess-capacity learning.” Traditionally, it was thought that the most effective way to learn was to simplify and forget extraneous details, but AI models have demonstrated that excessive complexity can actually enhance learning outcomes. This has profound implications as it contradicts a century of psychological research, which maintained that no system should retain irrelevant details given limitations on information storage. Co-authored by cognitive scientist Marina Dubova, the paper published in *Behavioral and Brain Sciences* proposes a new framework for understanding how humans might adopt a similar learning method. It delineates three levels of learning capacity—constrained, sufficient, and excess—where excess-capacity learning allows individuals to remember not just pertinent information but also seemingly irrelevant details. This could lead to superior generalizations and predictions based on expanded past experiences. By inviting interdisciplinary commentary, Dubova and her co-author aim to ignite a broader investigation into this uncharted territory of cognition, encouraging contributions from various fields including biologists and computer scientists.