"Understanding Concepts: An Essential Aspect of Robust Intelligence" PI: Patrick H. Winston MIT
ABSTRACT
This project will investigate foundational issues about the nature and representation of concepts, a critical prerequisite for building robust intelligence systems. In this project, which will be carried out in the Genesis Group of the MIT CSAIL laboratory, a concept is a complex, cross-modal model that crystallizes out of experience. As a step toward learning how to use experience, real and surrogate, to build such models, this project will collect and devise a rich collection of cross-linked, understanding-oriented experts, each specialized to a particular representation of the world, all backed by both individual associative memories and by associative memories that span multiple representations. The memories will be populated from a stream of sentences and phrases, thereby accumulating a humanlike capacity to associate states and actions that appear in one modality with those that appear in other modalities. This project will distill and test important aspects of concept representation by building a system called the Gauntlet System in which sentences and phrases stream past a line of representation-centered experts, each of which can either interpret an element of the data in its own terms or ignore it and pass it along to other downstream experts for their consideration. The project plans to build an on-line library of representation summaries for use by other researchers interested in concept formation, robust systems and Artificial Intelligence.