Semantic anchors facilitate task encoding in continual learning

• DOI: 10.31234/osf.io/wy7c4_v1
• Publication Date: 2025-03-08T09:52:33Z
• Authors: Mina Habibi; Pieter Verbeke; Mehdi Senoussi; Senne Braem

Humans are remarkably efficient at learning new tasks, in large part by relying on the integration of previously learned knowledge. However, research task typically focuses abstract rules minimalist stimuli, to study behavior independent history that humans come equipped with (i.e., semantic knowledge). In contrast, several theories suggest use knowledge and labels may help information. Here, we tested whether providing existing, semantically rich response allowed for more robust encoding less (catastrophic) forgetting interference. Our results show resulted (Experiment 1), both when using pictorial symbols or words as 2). Using artificial recurrent neural networks fitted behavior, conditions separated representations during learning. Finally, a subsequent value-based decision-making reinforcement modeling 3), demonstrate how embedding novel stimuli rich, representations, further efficient, feature-specific processing Together, our findings benefit learning, thereby offering important insights into why excel continual susceptible catastrophic compared most agents.