Reproduce “Y-Autoencoders: disentangling latent representations via sequential-encoding” on OpenResearch

About this paper

Massimiliano Patacchiola, Patrick Fox-Roberts, Edward Rosten

The Y-Autoencoder, developed by the University of Edinburgh and Snap Inc., learns disentangled latent representations in generative models through a new sequential-encoding training methodology for conditional autoencoders. This approach achieves superior disentanglement performance on tasks like digit manipulation, 3D object pose generation, and facial attribute transfer, while maintaining the training stability characteristic of traditional autoencoders.

Key insights

• A dual-branch sequential-encoding training procedure, rather than adversarial training, can enforce disentanglement in autoencoders by using the same implicit code with varying explicit labels.
• Carefully constructed loss functions, particularly an implicit loss that penalizes changes in the implicit code despite explicit label manipulations, are crucial for separating factors of variation.
• The Y-shaped structure in decoding, combined with re-encoding, allows the model to learn explicit (controllable) and implicit (style/texture) latent representations independently.

Summary by alphaXiv.