VAE Approximation Error
The importance of Variational Autoencoders reaches far beyond standalone gen- erative models — the approach is also used for learning latent representations and can be generalized to semi-supervised learning. This requires a thorough analy- sis of their commonly known shortcomings: posterior collapse and approximation errors. This paper analyzes VAE approximation errors caused by the combination of the ELBO objective and encoder models from conditional exponential families, including, but not limited to, commonly used conditionally independent discrete and continuous models. We characterize subclasses of generative models consis- tent with these encoder families. We show that the ELBO optimizer is pulled away from the likelihood optimizer towards the consistent subset and study this effect experimentally. Importantly, this subset can not be enlarged, and the respective error cannot be decreased, by considering deeper encoder/decoder networks. (Shekhovtsov et al., 2022)
Papers
- Shekhovtsov, A., Schlesinger, D., & Flach, B. (2022). VAE Approximation Error: ELBO and Exponential Families. International Conference on Learning Representations. https://openreview.net/forum?id=OIs3SxU5Ynl