Reproduce “Compositional learning of functions in humans and machines” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Yanli Zhou, Brenden M. Lake, Adina Williams

This research explores how humans and neural networks learn to compose functions with complex interactions in a visual domain, using a novel 'car assembly game'. It demonstrates that meta-learning-trained Transformers, when fine-tuned with human behavioral data, can closely match human generalization and error patterns in zero-shot compositional tasks.

Key insights

  • Humans exhibit robust compositional generalization across all tested complex function interaction types, achieving high accuracy without strong biases suggested by some linguistic theories.
  • AI models, specifically Transformers trained with MLC, can achieve near-human levels of accuracy in a challenging visual compositional task involving intricate function interactions.
  • Fine-tuning AI models with human behavioral data and synthetic noise effectively aligns their generalization patterns and error profiles closer to those observed in human participants.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.