Grounding the Unobservable in the Observable: The Role and Representation of Hidden State in Concept Formation and Refinement

Learning Grounded Representations

Volume

Papers from the 2001 AAAI Spring Symposium

Contents

One of the great mysteries of human cognition is how we learn to discover meaningful and useful categories and concepts about the world based on the data flowing from our sensors. Why do very young children acquire concepts like support and animate (Leslie 1988) rather than between three and six feet wide or blue with red and green dots? One answer to this question is that categories are created, refined and maintained to support accurate prediction. Knowing that an entity is animate is generally much more useful for the purpose of predicting how it will behave than knowing that it is blue with red and green dots. The idea of using predictability, or a lack thereof, as the driving force behind the creation and refinement of knowledge structures has been applied in a variety of contexts. Drescher (1991) and Shen (1993) used uncertainty in tion outcomes to trigger refinement of action models, and McCallum (1995) and Whitehead and Ballard (1991) uncertainty in predicted reward in a reinforcement learning setting to refine action policies. Virtually all of the work in this vein is based on two key assumptions. First, an assumption is made that the world is in priciple deterministic; that given enough knowledge, outcomes can be predicted with certainty. Given this, an agent’s failure to predict implies that it is either missing information or incorrectly representing the information that it has. Second, it is assumed that knowledge structures sufficient for the task can be created by combining raw perceptual information in various ways. That is, everything the agent needs to make accurate predictions is available in its percepts, and the problem facing the agent is to find the right combination of elements of its perceptual data for this task. (See (Drescher 1991) for an early and notable exception.) Our position is that the first of these assumptions represents an exceedingly useful mechanism for driving unsupervised concept acquisition, whereas blind adherence to the second makes it difficult or impossible to discover some of the most fundamental concepts.