One-Trial Learning through Explanatory Generalization

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 236 papers53 conceptsfrozen 2026-10-04

Most machine learning models are "data-hungry," requiring thousands of examples to identify a pattern. Humans, however, possess a remarkable ability for one-trial learning: seeing a single instance of a new concept—like a kidnapping or a new physics problem—and immediately grasping the underlying rule.

The research direction of Explanatory Generalization (often termed Explanation-Based Learning or EBL) posits that this efficiency is possible because we don't just look for statistical correlations; we construct a causal explanation for why an example works. By grounding a single observation in existing background knowledge, a system can strip away irrelevant details and extract a reusable template, or "schema," for the future.

The Intuition: Why vs. What

Traditional inductive learning is a process of "finding the common thread" across many examples. If you see ten Italians and they all speak Italian, you might induce that "Italians speak Italian." But as Stuart Russell 0001 points out in Preliminary Steps Toward the Automation of Induction, if you meet one Italian named Giuseppe, you don't generalize that all Italians are named Giuseppe.

The difference lies in our prior causal theories. We have a "determination" rule—a logical structure explained by Todd R. Davies and Stuart Russell 0001 in A Logical Approach to Reasoning by Analogy—that tells us nationality determines language, but nationality does not determine a first name. Explanatory generalization uses this background knowledge to justify why a single instance is representative of a whole class.

The Mechanism: From Trace to Template

How does a machine turn a specific event into a general rule? The process, pioneered by Gerald DeJong in Generalizations Based on Explanations, follows three steps:

  1. Explanation: The system uses its existing knowledge to "prove" how the observed goal was achieved. For example, in the GENESIS system described by Raymond J. Mooney and Gerald DeJong in Learning Schemata for Natural Language Processing, the system might explain a kidnapping by realizing it is a "steal" event used to trigger a "bargain" event.
  2. Generalization: The system looks at the logical trace of that explanation and identifies which parts were essential. It replaces specific constants (like "John" or "The Mona Lisa") with variables (like "Person X" or "Object Y").
  3. Schema Acquisition: This generalized logical structure is saved as a new "schema." The next time the system sees a similar sequence, it doesn't have to reason from first principles; it simply recognizes the schema.

This approach transforms learning from an empirical problem (counting occurrences) into an analytical one (understanding logic). As Paul O'Rorke argues in Generalization for Explanation-Based Schema Acquisition, this allows an "apprentice" system to learn complex heuristics just by observing a single expert demonstration.

Learning from Failure

Explanatory generalization is equally powerful when things go wrong. In Learning to Anticipate and Avoid Planning Problems through the Explanation of Failures, the CHEF system demonstrates that a single "kitchen nightmare" can be generalized into a broad rule. By analyzing the causal chain of a failed recipe, the system identifies the class of problematic features—such as "liquids that boil over"—rather than just remembering that specific pot of milk.

Similarly, Michael J. Pazzani showed in Refining the Knowledge Base of a Diagnostic Expert System that when a heuristic fails, a system can use a deep physical model to deduce a more accurate rule. This "failure-driven learning" ensures the system doesn't just memorize errors but understands the underlying mechanism of the mistake.

Modern Applications: Compositional and Robotic Learning

While the roots of this idea are symbolic, the principle of "learning by inverting a causal process" has seen a modern resurgence. Brenden M. Lake, Joshua B. Tenenbaum, and Ruslan Salakhutdinov demonstrated in One-shot learning by inverting a compositional causal process that modeling visual concepts as "compositional programs" allows machines to match human-level accuracy in one-shot character recognition.

In robotics and task automation, this manifests as extracting reusable "routines" from single demonstrations. Systems like RAPL, developed by Zelin Zhao and colleagues in Augmenting Policy Learning with Routines Discovered from a Single Demonstration, allow agents to decompose a single expert performance into high-level primitives. This mirrors the work of Aaron Mininger and John E. Laird in Interactively Learning a Blend of Goal-Based and Procedural Tasks, where agents use symbolic explanations to immediately apply learned task structures to entirely new environments.

Summary

The power of explanatory generalization lies in its sample efficiency. By shifting the burden from the quantity of data to the quality of background knowledge, these systems achieve a form of intelligence that is robust, transparent, and capable of mastering complex concepts from the very first encounter.

Papers in this version

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 42 papers
  2. Version 1 · 2026-10-04 · 36 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_one_trial_learning_b74041a1,
  title = {one-trial learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/one-trial-learning/rdr_b74041a16c2e476c8305439476735122},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_one_trial_learning_through_explanatory_g_b74041a1_v1,
  title = {One-Trial Learning through Explanatory Generalization},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/one-trial-learning-through-explanatory-generalization/rdr_b74041a16c2e476c8305439476735122/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}