Reproduce “Unsupervised Object Representation Learning using Translation and Rotation Group Equivariant VAE” on OpenResearch

About this paper

Alireza Nasiri, Tristan Bepler

TARGET-VAE, a variational autoencoder, learns semantic representations invariant to object translation and rotation by integrating group equivariance into both its encoder and decoder. This framework significantly improves unsupervised clustering accuracy on MNIST datasets, achieving up to 77.7%, and accurately infers pose parameters, proving effective for analyzing complex scientific imaging data like cryo-electron micrographs.

Key insights

• Integrating group equivariance into both the inference (encoder) and generative (decoder) networks is crucial for robustly disentangling semantic representations from spatial transformations in an unsupervised manner.
• The use of a mixture distribution over discrete rotation components, coupled with an explicit offset in the continuous rotation posterior's mean, significantly improves the accuracy of unsupervised pose inference.
• Increasing the level of discrete rotation equivariance (e.g., from P4 to P16) in the group convolutions consistently enhances both rotation correlation and the quality of disentangled semantic representations.

Summary by alphaXiv.