General Learning

IJCAI 1981

TL;DR

This paper introduces a concept learning system that incrementally builds and refines its knowledge base by treating concept descriptions as executable programs. It addresses the limitation of existing systems that learn concepts in a single session without leveraging prior knowledge, proposing a method where learned concepts augment the system's capabilities for future learning tasks.

What Problem Does This Paper Address?

Traditional concept learning systems often operate under two key assumptions: (1) all necessary information to learn a concept is present in the provided examples, and (2) all examples are available simultaneously for a single-session learning process. This approach overlooks the valuable aspect of learning systems that can improve their performance over time by incorporating previously acquired knowledge. The paper aims to bridge this gap by developing a system that can learn concepts incrementally, augmenting its knowledge base with each new concept learned, and using this expanded knowledge for future learning.

Methods

The core of the proposed system lies in its representation of concepts and its learning process. Concepts are represented using attribute-value pairs, and descriptions are formulated in a first-order predicate calculus-like language. A key innovation is treating these concept descriptions as executable programs. When the system needs to test a learned concept, it generates an instance of that concept and "executes" its description. This execution can be thought of as a form of theorem proving. The system learns by generalizing from positive and negative instances. When a generalization is made, it's tested by generating a new instance. If the trainer rejects this instance, the system analyzes the incorrect generalization to refine its concept description, ensuring it only allows valid instances. The system's knowledge base is augmented with each successfully learned concept, allowing it to reference previously learned concepts in new descriptions (e.g., "flat(X3)").

Datasets

Not discussed in the paper. The paper focuses on the methodology and provides illustrative examples rather than empirical evaluation on specific datasets.

Key Insights and Ideas

  • Concepts as Executable Programs: The idea of treating concept descriptions as programs that can be "executed" to generate instances or test hypotheses is a powerful mechanism for learning and verification.
  • Incremental Knowledge Augmentation: The system's ability to build upon its existing knowledge base by incorporating newly learned concepts is crucial for long-term learning and improved performance over time.
  • Learning through Experimentation and Refinement: The system learns by proposing hypotheses (generalizations) and testing them through simulated "experiments" (generating instances). Rejection of these instances by a trainer provides feedback for refining the concept description.
  • Generalization and Specialization: The paper demonstrates how the system can generalize statements (e.g., from specific shape/color to broader categories) and then specialize them again when necessary to avoid incorrect constructions, as seen with the "flat(X3)" constraint.

Key Results

The paper demonstrates the system's capability to learn concepts by example, specifically illustrating the learning of an "on-top-of" concept. It shows how the system can generalize from a specific instance and then refine its understanding based on feedback. The authors suggest that the system, although simple, is capable of learning complex concepts, including recursive ones like list reversal, once the "reverse" concept is learned.

Limitations and Future Work

The authors acknowledge that the described system is simple. They suggest that the learning process can be viewed as a form of automatic programming. Future work could involve exploring more complex concept representations and learning algorithms. The paper also implies that learning can be interrupted and resumed, suggesting a modular approach to concept acquisition.

Summary updated

BibTeX
@inproceedings{1981general,
  title = {{General Learning}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1981},
}
Figure