NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 25·8 min read
A mathematical theory of balancing relational generalization and memorization
TL;DR
Learning systems must balance learning general rules versus memorizing exceptions, but existing benchmarks don't test this ability. Researchers introduced transitive inference with exceptions—a task where models learn ordered relationships while handling rule violations—and analytically characterized kernel ridge regression behavior across different representations.
✦ Why It Matters
Engineers can predict when models will fail at balancing generalization and memorization by understanding representational geometry requirements.
Key Takeaways
- Test model generalization versus memorization using transitive inference with exceptions task paradigm.
- Representational geometry critically determines success; audit feature space structure before deployment.
- Kernel ridge regression theory predicts systematic failure modes; validate against your model architecture.
- Pretrained language models show predictable errors on exception handling; fine-tune carefully on edge cases.
Full Summary
Intelligent systems must learn generalizable rules while accommodating exceptions, yet few benchmarks directly test this capability. Cheng and Lippl introduced transitive inference with exceptions, a task paradigm where models learn transitive relations (if A>B and B>C, then A>C) while also memorizing exceptions to this rule.
They analytically characterized kernel ridge regression—a theoretically tractable neural network model—across diverse representations and task parameters. Unlike transitive inference without exceptions, successful performance here depends sensitively on representational geometry (how data is structured in the model's feature space).
The authors explain mechanistically why this task is harder and validate predictions using pretrained language models finetuned on ordered relations, finding models generalize correctly but make systematic errors matching theoretical predictions.