Insight under the learning lens - What spontaneous learning dynamics can teach us about insight

Abstract

Insight is widely considered as a distinctive facet of human cognition, closely linked to creativity and metacognitive reasoning. It manifests as sudden comprehension and marked improvements in performance. Despite over a century of research, the cognitive and neural underpinnings of insightful problem solving remain subjects of ongoing debate. This thesis investigates insight as a learning phenomenon in both humans and artificial neural networks, with a particular focus on the role of sleep as an incubator for insight and its relation to computational mechanisms underlying insight-like learning. Chapter 1 begins by tracing the historical trajectory of insight research, introducing experimental paradigms and related key concepts such as impasse, incubation, suddenness and affective components, neural correlates, and computational models. The second section of the chapter introduces the PSSST — a novel hidden rule insight paradigm designed to investigate the acquisition of both a hidden and explicit strategy with high temporal precision. This unique paradigm offers a fine grained temporal window into the learning dynamics that culminate in insight. In Chapter 2, the PSSST is employed to characterise insight in human participants based on three core attributes: suddenness, selectivity and delay. These behavioural characteristics allow for the comparative study of insight-like behaviour across biological and artificial intelligence systems, providing a foundation for understanding the computational mechanisms that drive insight. Chapter 3 focuses on this comparison between humans and artificial neural network learning dynamics. Simple neural networks, devoid of complex cognitive processes, exhibit learning dynamics that qualitatively and quantitatively align with human behavioural characteristics of insight. Further analysis of different network architectures and learning trajectories revealed that a regularised gating mechanism and noise added to gradient updates are crucial for eliciting insight-like behaviour. These mechanisms in combination enable the accumulation of “silent knowledge” — initially masked by regularised gating. These findings suggest a link between behavioural markers of insight and measurements of noise and regularisation and their neural counterparts in biological systems, thereby opening new avenues for investigating insight in the brain. Chapter 4 examines the relationship between insight and sleep, identifying a significant role for N2 sleep and aperiodic neural activity. Participants who reached N2, but not N1 sleep, showed an increased likelihood of gaining insight on the PSSST after a daytime nap, implying a role of deeper sleep for insight. EEG power spectrum analysis further demonstrated that the spectral slope best predicted insight, beyond sleep stages alone. These findings align with the theory that regularisation is linked to synaptic downscaling during deeper sleep, reinforcing the proposed connection between insight and regularisation in both computational and biological contexts.