Prospective Compression in Human Abstraction Learning
A new study on arXiv (2605.09985) investigates how humans learn abstractions in non-stationary environments, using the Pattern Builder Task where participants create geometric patterns with primitives and custom helpers. The research proposes that humans select abstractions prospectively to compress future tasks, unlike existing algorithms that rely on retrospective compression over static task distributions. Two experiments were conducted to test this hypothesis.
Key facts
- arXiv paper 2605.09985
- Studies prospective compression in human abstraction learning
- Uses Pattern Builder Task
- Participants create geometric patterns from primitives, transformations, and helpers
- Contrasts with retrospective compression algorithms
- Two experiments conducted
- Focus on non-stationary domains
- Hypothesis: humans target compression of future tasks
Entities
Institutions
- arXiv