Abstract
Traditional incremental learning systems often suffer from extreme conservatism, leading to confirmation bias and an inability to escape local maxima when faced with noisy or complex data. This paper introduces METAXA.3, a system capable of non-cumulative knowledge revision. By monitoring for theoretical "crises," METAXA.3 can abandon failing paradigms in favor of more robust theories structured around higher-level meta-concepts.
Introduction
During the 1980s, incremental learning research was dominated by conservative strategies. These systems aimed to minimize changes to existing knowledge when encountering new data. While this approach ensures stability, it frequently traps systems in "local maxima"—mediocre theories cluttered with endless exceptions. Werner Emde argues that for a system to truly learn, it must be capable of "scientific revolutions": radical restructurings of its knowledge base when current theories prove inadequate.
System Architecture
METAXA.3 is structured to support both standard incremental updates and radical revisions. The architecture centers on a "TELL/ASK" interface that mediates interaction between the user and the core processing units.
The system comprises three primary modules: the Inference Engine, which handles facts and heuristics; the Induction Program, which generates hypotheses; and the Knowledge Revision module, which manages theoretical transitions. A key innovation in METAXA.3 is the use of Higher Concepts (meta-facts and meta-rules). The system does not merely search for simple correlations; it identifies structural patterns such as transitivity or constant ratios between properties.
Non-Cumulative Knowledge Revision
The system operates in two modes. In Cumulative Mode, it performs conservative adjustments. However, when a "crisis" is detected—defined by exceeding a threshold of contradictory data or discovering a powerful new meta-concept—it switches to Non-Cumulative Mode.
In this mode, the system employs heuristic-driven shifts. For example, if the system observes a constant ratio between weight and volume across multiple objects, a const_ratio heuristic may trigger the induction of a "density" concept. This allows the system to reorganize its entire factual database under a new paradigm, often "rehabilitating" data previously dismissed as noise.
Case Study: The Physics of Buoyancy
The efficacy of METAXA.3 was demonstrated using a qualitative physics domain. Initially, the system might hold a simple theory: "Small objects float; large objects sink."
- Conflict: Encountering a small metal needle (which sinks) and a small floating cable creates tension in the initial theory.
- Discovery: The induction program identifies a rule regarding the specific weight of materials.
- Paradigm Shift: The
const_ratioheuristic triggers, leading the system to realize that the ratio of weight to volume (density) is the governing principle. - Result: The system adopts a density-based theory. The "noisy" needle is no longer an exception but a core supporting fact of the new, more elegant model.
Discussion and Limitations
METAXA.3 represents an early AI implementation of Thomas Kuhn’s philosophy of scientific revolutions. By explicitly modeling the need for paradigm shifts, it avoids the stagnation typical of purely conservative learners. However, challenges remain in computationally comparing the "elegance" of competing theories and selecting appropriate heuristics for more abstract domains. Future work is required to optimize search depths for counterexamples in larger, more complex datasets.
Summary updated