NVAE- A Deep Hierarchical Variational Autoencoder

Explore the intricacies of deep hierarchical Variational Autoencoders (VAEs) in this comprehensive 34-minute video explanation. Delve into the engineering choices necessary for training VAEs that produce high-resolution images with global consistency and remarkable sharpness. Learn about the challenges of traditional VAEs, including training difficulties at high resolutions and instability in deep architectures. Discover the Nouveau VAE (NVAE) model, which utilizes depth-wise separable convolutions, batch normalization, and spectral regularization to achieve state-of-the-art results among non-autoregressive likelihood-based models. Gain insights into the hierarchical VAE decoder and encoder structures, output samples, and experimental results across various datasets. Understand the residual parameterization of Normal distributions and the importance of KL divergence from deltas in stabilizing training.

- Intro & Overview
- Variational Autoencoders
- Hierarchical VAE Decoder
- Output Samples
- Hierarchical VAE Encoder
- Engineering Decisions
- KL from Deltas
- Experimental Results
- Appendix
- Conclusion