Abstract

Turning regularities into categories is an important aspect of human cognition. We can make generalizations about new events and entities based on the categories we think they belong to. Structuring knowledge into categories also facilitate search and retrieval. Moreover, remembering specific instances of categories (e.g., the first day at a job) is crucial for how we process information. Similarly, artificial intelligence systems require the capacity to represent and reason about both categories and instances. In this talk, I describe two tasks inspired by experiments in developmental psychology for evaluating this capacity. The first task, novel noun generalization, examines whether our existing models can determine the correct level of a hierarchical taxonomy (e.g., dog or animal) a novel word refers to. The second task evaluates models' ability to represent different states of the world (i.e., position of an item). I discuss how current models perform on these tasks and what inductive biases can help models succeed.

Slides