Learning Physical Descriptions From Functional Definitions, Examples, and Precedents

Patrick H. Winston, Boris Katz, Thomas O. Binford, Michael R. LowryAAAI 1983 · OpenReview

Abstract

It is too hard to tell vision systems what things look like. It is easier to talk about purpose and what things are for. Consequently, we want vision systems to use functional descriptions to identify things, when necessary, and we want them to learn physical descriptions for themselves, when possible. This paper describes a theory that explains how to make such systems work, synthesizing ideas about learning from precedents with physical description modeling.

Introduction: The Function-to-Form Gap

In early computer vision research, a primary bottleneck was the sheer variety of physical forms a single object class could take. A "cup" might be made of various materials, possess different shapes, or lack a handle entirely. Describing every possible physical permutation is an intractable task.

The authors propose that it is more efficient to define objects by their function. A cup is functionally defined as a stable, liftable, open vessel. The core challenge then becomes the "function-to-form mapping": how can a system determine that a "flat bottom" satisfies the functional requirement of "stability"?

Methodology: Learning via Analogy and Precedents

The proposed system bridges this gap by using precedents—knowledge of other objects whose physical traits and functional successes are already known. When presented with a new physical example and a functional definition, the system performs an analogy to derive the necessary physical model.

Analogy through matching precedents to current problems.

The process involves five key steps:

  1. Functional Description: A definition (e.g., "A cup is a stable, liftable open-vessel") is translated into a semantic network.
  2. Physical Example: The system receives a geometric representation of a specific object.
  3. Description Enhancement: Non-visual properties like material or weight are added.
  4. Analogy via Precedents: The system matches parts of the new object to known precedents. For instance, it might use a "brick" to understand that a flat bottom causes stability, or a "suitcase" to understand how a handle enables liftability.
  5. Rule Generation: Once the causal links are established, the system creates a symbolic rule.

Rule Generation and Censors

A significant contribution of this work is the automated construction of symbolic if-then rules. These rules allow the system to recognize similar objects in the future without repeating the computationally expensive analogy process.

Structure of a generated symbolic rule.

Furthermore, the system employs "censors" to handle exceptions and conditional logic. For example, while a handle usually makes a cup liftable, a cup containing hot liquid might be ungraspable unless it is made of an insulating material like Styrofoam. These censors allow the system to reason about complex physical realities beyond simple geometric matching.

Key Results and Insights

The implemented system successfully demonstrated that functional definitions are more robust than physical ones. By understanding why a feature exists (e.g., a handle for graspability), the system can recognize a cup even if the handle's specific shape is novel. The research showed that "learning" in this context is not a statistical process but the automated construction of a causal structure that explains an object's design.

Limitations and Conclusion

Despite its success, the system relied heavily on manual knowledge engineering; humans had to hand-code the precedents and functional definitions. Additionally, the integration with real-world vision systems was limited at the time, often requiring English descriptions of geometry rather than raw visual input. However, by combining functional reasoning with analogical learning, this framework provided a path for vision systems to "understand" the objects they see, moving AI closer to a paradigm where form follows function.

Summary updated

Abstract

It is too hard to tell vision systems what things look like. It is easier to talk about purpose and what things are for. Consequently, we want vision systems to use functional descriptions to identify things, when necessary, and we want them to learn physical descriptions for themselves, when possible. This paper describes a theory that explains how to make such systems work. The theory is a synthesis of two sets of ideas: ideas about learning from precedents and exercises developed at MIT and ideas about physical description developed at Stanford. The strength of the synthesis is illustrated byway of representative experiments. All of these experiments have been performed with an implemented system.
BibTeX
@inproceedings{winston1983learning,
  author = {Patrick H. Winston and Boris Katz and Thomas O. Binford and Michael R. Lowry},
  title = {{Learning Physical Descriptions From Functional Definitions, Examples, and Precedents}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
  url = {https://openreview.net/forum?id=SkZJqpeOZH},
}
Figure