Object Representations for Learning and Reasoning

Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS)

December 11, 2020, Virtual Workshop

Join via the livestream 🎥 · RocketChat · @ORLR_Workshop · #ORLR2020 · Join our community Slack!

---

Abstract

Research over the last few decades has shed considerable light on how infants represent and reason about objects in simple physical events, such as occlusion, containment, and support events. My talk summarizes key findings from this research. First, representations of objects in physical events are initially very sparse and lacking in featural detail; nevertheless, they are still sufficient, when interpreted by infants’ core physical knowledge, to allow infants to correctly predict the outcomes of many physical events. Second, object representations become richer and more detailed as infants form event categories and, for each category, identify features that are causally relevant for predicting outcomes. Third, once a feature has been identified as relevant to an event category, it is represented for any event in the category, thereby ensuring powerful generalization. Fourth, features identified in one event category are not generalized to other event categories, even when equally relevant, resulting at times in striking discrepancies in infants’ responses to similar events from different categories. Fifth, one of the learning mechanisms by which infants identify features is explanation-based learning; because this mechanism uses analytical as well as empirical evidence, it is highly efficient and allows infants to acquire new featural rules with few exemplars. Finally, the cognitive architecture that underlies early representations of objects in physical reasoning includes at least two systems, an object-file and a physical-reasoning system, each with a different role; understanding how each system represents objects, how these representations change with experience, and how the two systems exchange information as events unfold helps explain the findings described above. Together, these findings make clear that object representations in infancy are far from monolithic: They are both system-specific (the same object may be represented differently in different systems) and event-specific (the same object may be represented differently in different events), and they become gradually richer with experience.