Learning by Re-Expressing Concepts for Efficient Recognition

Richard M. KellerAAAI 1983

Abstract

Traditional machine learning research has focused heavily on the induction of new concepts from examples. However, a significant challenge in intelligent systems is the "operationalization" of existing knowledge—transforming accurate but computationally expensive functional definitions into efficient, structural descriptions. This paper presents a framework for concept reformulation through a library of transformation techniques, demonstrating how a system can "learn" to recognize instances faster by re-expressing its internal knowledge.

Introduction: The Operationalization Problem

In the early 1980s, Richard Keller identified a critical gap in AI: the distinction between knowing a concept and being able to use it efficiently. While a functional definition (e.g., an arch is a structure that supports weight over an opening) is theoretically complete, verifying it in a real-world scenario might require intensive physics simulations. Conversely, a structural definition (e.g., three specific blocks in a particular arrangement) is easy to check but often missing from a system's initial knowledge base.

Keller defines this as a transition from an Inefficient Recognizer Language (IRL) to an Efficient Recognizer Language (ERL). The goal of learning, in this context, is not to discover new truths, but to compile abstract knowledge into fast, reactive heuristics.

Methodology: Transformation Techniques

The core of the proposed framework is a library of Operationalization Techniques. These are formal transformation rules applied to predicate calculus expressions to move them toward the ERL. These techniques are categorized based on how they affect the semantic scope of the concept being transformed.

Table of transformation rules categorizing techniques into concept-preserving, specializing, and generalizing transforms.

As shown in the table above, these transforms include:

  • Concept-Preserving Transforms: Such as Definitional Expansion, which replaces a term with its logical equivalent to expose more primitive features.
  • Concept-Specializing Transforms: Such as Disjunct Selection, which narrows a concept to a specific, easily-verifiable case, trading off generality for speed.
  • Concept-Generalizing Transforms: Such as Conjunct Deletion, which removes hard-to-verify requirements, potentially introducing false positives but significantly reducing computational cost.

Case Study: The LEX System

The theory was evaluated using the LEX system, a program designed to solve integral calculus problems. LEX initially possesses a high-level "Minimum Cost Criterion" for selecting mathematical operators (e.g., Integration by Parts), but this criterion is non-operational because it requires searching the entire problem space to verify.

Through a sequence of six transformation steps, the system can derive a structural pattern-matching rule from this abstract goal. For instance, it can learn that if an integral matches the pattern $\int \sin(x) dx$, a specific operator should be applied immediately.

The step-by-step operationalization of a rule in the LEX system, moving from abstract goals to concrete pattern matching.

This process, illustrated in Figure IV-2, demonstrates how techniques like Constraint Propagation and Redundancy Elimination are used to "compile" the high-level logic into a form that the system can recognize instantly during problem-solving.

Key Insights and Discussion

Keller’s work highlights that operationality is relative. A concept's utility is tied to the computational resources of the recognizer and the specific goals of the system. Learning is viewed as a heuristic search through a space of possible recognizers, guided by domain-specific bridge knowledge (e.g., physics or mathematics) that justifies why certain structural features can substitute for functional ones.

Furthermore, the system must maintain a "Knowledge Justification"—a logical trail explaining why certain features (like the color of a block in an arch) were deemed irrelevant and dropped during the specialization or generalization process.

Conclusion and Future Work

While the paper provides a robust theoretical foundation for knowledge compilation, it acknowledges significant challenges in search control. The space of possible transformations is vast, and determining the optimal sequence of rules remains a difficult problem. Future work in this area points toward more automated, general-purpose operationalizers that can dynamically adapt their recognition strategies based on the constraints of their environment.

Summary updated

BibTeX
@inproceedings{keller1983learning,
  author = {Richard M. Keller},
  title = {{Learning by Re-Expressing Concepts for Efficient Recognition}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1983},
}

Loading related content...

Figure