Spontaneous emergence of context-dependent statistical learning in humans and neural networks

Humans readily extract statistical regularities from experience, yet natural environments require flexible adaptation when associative structures shift across changing contexts, often without warning. Across two experiments, we show that humans can incidentally learn overlapping and conflicting visual associations even when contexts dynamically alternate and remain unsignaled or only minimally cued. To probe the computational mechanisms supporting this adaptive capacity, we trained recurrent neural networks with gated recurrent units on the same statistical learning task without providing any explicit context information. These models spontaneously developed distributed internal representations that robustly separated conflicting associations and supported rapid adaptation to latent context shifts. Critically, we show that these distributed representations, strongly shaped by the model's initial weight configuration, played a key role in preventing catastrophic interference between contexts. Together, these behavioral and computational results significantly advance our understanding of how humans and artificial systems can successfully learn and flexibly retrieve context-dependent associations under challenging conditions.