An Algorithm for Learning Shape and Appearance Models without Annotations.

•A framework for automatically learning shape and appearance models without manual annotations.•Designed to run within a distributed privacy preserving framework.•When used as a pattern recognition approach, can give competitive classification accuracies for MNIST - particularly for small numbers of training examples.•Can handle missing data in the images.•Tested the model with 1900 brain scans and found that its latent variables can be used as features for pattern recognition.