Learning from the Schema Learning System

A major problem that confronts developers of knowledge-based vision systems is how the information in the knowledge base (both objectspecific knowledge and the control strategies for applying the knowledge) is acquired. Hand construction of knowledge bases is an onerous, time-consumlng and errorful process, and is inadequate for systems requiring more than a few object models, especially if the domain within which the system operates changes. Over the past four years we have addressed this problem by developing the Schema Learning System (SLS) to learn strategies for applying objectspecific knowledge in complex domains. The goal is build a system that can learn (under supervision) to recognize a new object or object class without any direct human intervention. This paper will not describe SLS in any detail; that has been done elsewhere. (See Draper et al. [9] for a preliminary presentation, or Draper [I0] for a more recent and complete description). Instead, this paper summarizes some of our conclusions from four years of machine learning and computer vision research, emphasizing our representation of recognition strategies, our syntactic approach to integrating visual modules, and the implications of SLS for goal-directed, as opposed to reconstructionist, theories of vision.