Reproduce “Interaction Asymmetry: A General Principle for Learning Composable Abstractions” 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
Jack Brady, Julius von Kügelgen, Sébastien Lachapelle, Simon Buchholz, Thomas Kipf, Wieland Brendel
Researchers define "Interaction Asymmetry," a general principle formalizing how concepts are structured by their internal versus external interaction complexity, and prove it enables learning disentangled representations and compositional generalization. Their method, an attention-regularized Transformer-VAE, empirically disentangles objects and models complex interactions in visual scenes more effectively than previous approaches.
Key insights
- Interaction asymmetry, formally defined by higher-order derivative conditions, provably enables disentanglement of concepts into independent latent slots and compositional generalization to unseen concept combinations.
- The theoretical framework unifies and extends prior work on disentanglement for non-interacting (n=0) and additively interacting (n=1) objects to higher orders (n=2), allowing for more complex polynomial interactions between concepts.
- The inherent complexity of interactions within a concept must exceed the complexity of interactions between distinct concepts for identifiability and learnability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.