Reproduce “Disentangling Factors of Variation with Cycle-Consistent Variational Auto-Encoders” on OpenResearch

About this paper

Ananya Harsh Jha, Saket Anand, Maneesh Singh, V. S. R. Veeravasarapu

This research introduces Cycle-Consistent Variational Auto-Encoders (Cycle-Consistent VAEs), a non-adversarial architecture that achieves disentangled representations using only pairwise similarity labels. The method successfully separates specified and unspecified factors of variation, demonstrated by significantly reducing classification accuracy on the unspecified latent space to 17.72% on MNIST, while maintaining high accuracy on specified factors and enabling controlled image generation.

Key insights

• Cycle-consistency, applied through a novel reverse cycle loss, can effectively drive the encoder to explicitly disentangle latent representations, preventing information leakage of specified factors into the unspecified latent space.
• Effective disentanglement is achievable with weak supervision (pairwise similarity labels), significantly reducing annotation burden compared to methods requiring full class labels or triplets.
• The proposed non-adversarial architecture provides robustness against degenerate solutions and insensitivity to latent space dimensionality, addressing key limitations of prior disentanglement methods.

Summary by alphaXiv.