Proceedings of the Annual Meeting of the Cognitive Science Society

From an early age, humans are capable of learning about theirsocial environment, making predictions of how other agentswill operate and decisions about how they themselves will in-teract. In this work, we address the problem of formalizing thelearning principles underlying these abilities. We construct a cu-rious neural agent that can efficiently learn predictive models ofsocial environments that are rich with external agents inspiredby real-world animate behaviors such as peekaboo, chasing,and mimicry. Our curious neural agent consists of a controllerdriven by γ-Progress, a scalable and effective curiosity signal,and a disentangled world model that allocates separate networksfor interdependent components of the world. We show that ourdisentangled curiosity-driven agent achieves higher learning ef-ficiency and prediction performance than strong baselines. Cru-cially, we find that a preference for animate attention emergesnaturally in our model, and is a key driver of performance. Fi-nally we discuss future directions including applications of ourframework to modeling human behavior and designing earlyindicators for developmental variability.