Unsupervised Learning of Group Invariant and Equivariant Representations

Explore unsupervised learning of group invariant and equivariant representations in this comprehensive 1 hour 32 minute video from the Learning on Graphs and Geometry Reading Group. Delve into the extension of group invariant and equivariant representation learning to unsupervised deep learning, focusing on a general learning strategy using an encoder-decoder framework. Discover how networks can learn to encode and decode data to and from group-invariant representations by predicting appropriate group actions for input-output pose alignment. Examine the necessary conditions for equivariant encoders and their construction for various groups, including rotations, translations, and permutations. Follow along as the presenters discuss motivations behind symmetry in data, define representations and equivariance, and provide practical examples for different groups. Gain insights into the application of these concepts to graphs and observe results for E(3) and Sn symmetry. Participate in two Q&A sessions to deepen your understanding of this cutting-edge approach in machine learning.

- Intro
- Motivation: Symmetry in Data
- Defining Representations and Equivariance
- Unsupervised Invariant Representation Learning
- Q+A
- Some Practical Examples for Groups
- Examples and Results for Graphs
- Results: E3 and Sn Symmetry
- Q+A