“Learning Structural Descriptions from Examples” was Patrick Winston’s MIT doctoral thesis, published in September 1970 as MIT AI Laboratory Technical Report 231 (AITR-231) and archived in MIT’s DSpace repository. It was a landmark in understanding how a machine might learn concepts that involve complex structural relationships, rather than simple properties, and it worked in the blocks world of toy bricks that was the era’s favorite testbed.
Winston’s program learned structural concepts such as “arch” by building and refining an abstract description of a scene. Shown a positive example of an arch, it formed a tentative model: two upright blocks supporting a third lying across the top. The pivotal idea was the use of the near miss, a carefully chosen counterexample that fails to be an arch in exactly one significant way. Shown a structure identical to an arch except that the two uprights touch each other, the program could deduce that the uprights must not touch and could tighten its model accordingly. By alternating positive examples with such instructive near misses, the program converged on a structural description that captured the concept, including which relations were required and which were merely incidental.
This was one of the first convincing demonstrations of learning from examples in AI, and the near-miss strategy, learning efficiently from examples that differ in just one telling respect, became a recurring theme in machine learning and in thinking about how to teach concepts.
Why a business reader should care: Winston’s program is an early answer to a question every modern machine-learning team faces, namely how a system can generalize a concept from a handful of well-chosen examples, and its insight that a single sharp counterexample can teach more than many similar positives still rings true.