An Integrated Systems Approach to Explanation-Based Conceptual Change

Scott E. Friedman, Kenneth D. ForbusAAAI 2010 · DOI

Human intelligence is characterized by more than just the accumulation of facts; it involves conceptual change, the process by which we radically restructure our existing knowledge. This phenomenon is most visible in childhood development, such as when a child transitions from believing "heavy things fall faster" to understanding the principles of universal gravitation. This paper, originally presented at AAAI-10, explores an integrated AI system designed to replicate this fundamental reorganization of ideas.

The Challenge: Beyond Data Accumulation

In traditional AI, learning is often framed as statistical pattern matching or simple data accumulation. However, human learning involves a fundamental reorganization of internal logic. For instance, a child's understanding of "force" typically evolves from an internal property of heavy objects to an "acquired" property during motion, and finally to an interaction between objects.

Existing AI models often struggle with this because they are brittle; they cannot easily "unlearn" or rewrite their internal logic when faced with anomalies. Furthermore, most models rely on pre-processed symbolic data. This research addresses these gaps by creating a system that interprets raw-ish visual input—sketches—and performs high-level theory revision to evolve its understanding of physics.

System Architecture: The Tiered Network

The authors built their model on the Companions cognitive architecture, integrating several sophisticated components to handle different aspects of reasoning and perception. A central feature of this architecture is the Tiered Network, which organizes knowledge into three distinct layers.

The tiered network architecture used for explanation-based conceptual change.

As shown in the architecture diagram, the system maintains:

  • Explanations (E): The current "story" or model of an event.
  • TMS Justifications (T): A Truth Maintenance System that records the logic behind beliefs, allowing the system to identify which parts of a theory need revision when contradictions arise.
  • Conceptual Knowledge (C): The underlying qualitative physics models.

Other key components include CogSketch, which interprets spatial relationships in hand-drawn sketches, and Qualitative Process (QP) Theory, which represents physical change through qualitative relationships (e.g., "if force increases, velocity increases") rather than complex differential equations.

Methodology: Learning from Sketches

The system was trained using 10 hand-drawn comic strips comprising 58 frames. These strips depicted various physical scenarios, such as a person pushing a wheelbarrow or a foot kicking a ball. The system uses the Structure-Mapping Engine (SME) for analogical reasoning, allowing it to compare new experiences to past events and "borrow" explanations from similar scenarios.

A sequence of states from a hand-drawn comic strip used as training input.

When the system encounters an anomaly—a physical behavior it cannot explain—it triggers conceptual change. Following a "simplicity heuristic," the system prefers making small tweaks to its models before attempting radical revisions. One of the most human-like features of the model is retroactive explanation: when a new model is learned, the system re-evaluates past memories to see if they can be better explained by the new knowledge.

Results: Replicating Human Developmental Stages

In simulations, the AI followed a developmental trajectory strikingly similar to human children across four distinct stages:

  1. Initial State: No concept of force.
  2. Internal Force: Force is perceived as an internal property of objects, often tied to size.
  3. Acquired Force: The system concludes objects "get" force when they are moved.
  4. Push/Pull: The system settles on a model where force is an external influence exerted by an agent.

While the AI progressed through these stages much faster than a human (10 trials versus years of schooling), the authors attribute this to the "clean" nature of the sketched input compared to the noisy data children encounter in the real world.

Conclusion and Future Directions

This research demonstrates that an integrated systems approach—combining qualitative reasoning, analogy, and truth maintenance—can successfully model the complex process of conceptual change. Future work aims to address the challenges of noisy real-world perception and integrate natural language processing to allow the system to learn from social instruction as well as direct experience. Additionally, the authors hope to model "knowledge shields," the human tendency to ignore anomalies to protect existing beliefs, making the AI's learning trajectory even more human-like.

Summary updated

BibTeX
@inproceedings{friedman2010integrated,
  author = {Scott E. Friedman and Kenneth D. Forbus},
  title = {{An Integrated Systems Approach to Explanation-Based Conceptual Change}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2010},
  doi = {10.1609/aaai.v24i1.7572},
}
Figure