HAG Model Combines Rate-Distortion Theory with Program Induction for Human-Like Sequence Learning
New model explains how prior knowledge shapes future abstractions under cognitive limits.
A team of cognitive scientists led by Hanqi Zhou developed HAG, a hierarchical model that unifies rate-distortion theory with program induction to explain how humans learn sequences under resource constraints. The model posits that prior knowledge shapes which structures are cheap to encode through separate local (within-task) and global (across-task) libraries, governed by memory and computation limits. In simulations, HAG achieved better rate-distortion trade-offs and stronger generalization than fixed grammars or shallow chunking methods.
In an online melodic sequence learning experiment, participants' recall errors reflected systematic simplifications, and reaction times spiked at boundaries inferred by HAG. Trial-by-trial fits showed HAG's hierarchical libraries best explained individual differences in recall and out-of-sample continuation choices, outperforming all alternative models. These findings portray structured learning as path-dependent program induction under bounded resources, with implications for building AI that learns more like humans.
- HAG integrates rate-distortion theory with program induction using hierarchical local and global libraries.
- Outperforms fixed grammars and shallow chunking in simulations on rate-distortion trade-offs and generalization.
- Human melodic sequence experiment: recall errors and reaction times aligned with HAG's inferred program boundaries.
Why It Matters
This model provides a computational foundation for AI systems that leverage prior experience to learn efficiently under constraints.