Abstract:

Humans excel at learning from the ever-present social stimuli in their environment.

To understand this cognitive capacity, computational models of social learning often borrow from the longstanding tradition of computationally modelling individual learning; reinforcement learning models can be augmented with social information to model imitation behaviour, and planning models can be inverted to model inference behaviour.

While a fruitful approach, this influence comes at a cost: social cognition research commonly adapts experimental designs and analysis methods from individual learning, treating social information as little more than a reward signal. This strips social cognition of the richness of stimuli we experience in even the most banal everyday situations, and risks systematically underestimating of our social learning capabilities.

In this talk, I will present results from three projects which span key domains of individual learning, applying them in contexts that are more naturalistic than the bandit tasks commonly used in the field to assess both individual and social learning: generalization, heuristics, and resource-rational strategy arbitration.

Across these three domains, I show that social learning is characterized by distinct behavioural patterns when testing individual learning principles in settings that preserve some of the richness of our social environments. Participants often performed better in social contexts, underscoring the importance of investigating human intelligence in the settings we evolved to excel in.