Reproduce “Compositional Generative Modeling: A Single Model is Not All You Need” 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

Yilun Du, Leslie Kaelbling

Compositional generative modeling constructs complex AI systems by combining simpler, specialized generative models, offering a more data-efficient, generalizable, and reprogrammable alternative to increasingly large, monolithic architectures. This approach demonstrates superior data efficiency, enhanced generalization to unseen tasks, and the ability to discover underlying compositional structures, challenging the notion that 'a single model is all you need'.

Key insights

  • Compositionality drastically reduces sample complexity compared to learning a single joint distribution, enhancing data efficiency.
  • Pre-trained generative components can be re-purposed and combined to solve novel tasks zero-shot without explicit retraining.
  • Meaningful compositional structures and components can be discovered unsupervisedly from raw data, leading to more interpretable and manipulable representations.

Summary by alphaXiv.

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