Summary of Interaction Asymmetry: a General Principle For Learning Composable Abstractions, by Jack Brady et al.
Interaction Asymmetry: A General Principle for Learning Composable Abstractions
by Jack Brady, Julius von Kügelgen, Sébastien Lachapelle, Simon Buchholz, Thomas Kipf, Wieland Brendel
First submitted to arxiv on: 12 Nov 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Computer Vision and Pattern Recognition (cs.CV)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary This paper proposes a principle called interaction asymmetry, which states that parts of the same concept have more complex interactions than parts of different concepts. The authors formalize this using block diagonality conditions on the higher-order derivatives of a generator mapping concepts to observed data. They prove that interaction asymmetry enables both disentanglement and compositional generalization, unifying recent theoretical results for learning object concepts. The authors provide results for up to n=2, extending prior works to more flexible generator functions. Practically, their theory suggests penalizing latent capacity and concept interactions during decoding. They propose a Transformer-based VAE with a novel regularizer on attention weights to implement these criteria. On synthetic image datasets, the model achieves comparable object disentanglement to existing models using explicit object-centric priors. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper is about understanding how we can learn and mix concepts together in new ways. The authors propose a rule called interaction asymmetry, which says that parts of the same concept are more connected than parts of different concepts. They show that this rule helps us understand how to separate and combine concepts, making it easier for computers to learn and generalize. The authors also suggest a new way to train autoencoders, using attention weights to control how much connections between concepts are learned. They test their idea on synthetic images and find that it works well. |
Keywords
» Artificial intelligence » Attention » Generalization » Transformer
