Lots of papers have been trying to address the problem of posterior collapse with VAEs. Due to huge number of publications, I thought it is intersting to have a list of related papers.
| Paper | Implementation | Publication Year | Citation |
|---|---|---|---|
| Ladder variational autoencoders | Yes | 2016 | 200+ |
| Fixing a broken ELBO | No | 2017 | 70+ |
| Neural discrete representation learning | Third Party | 2017 | 128+ |
| Tackling Over-pruning in Variational Autoencoders | No | 2017 | 20+ |
| Filtering Variational Objectives | Yes | 2017 | 60+ |
| Auxiliary Guided Autoregressive Variational Autoencoders | Third Party | 2017 | 10- |
| VAE with a VampPrior | Yes | 2017 | 70+ |
| Z-Forcing: Training Stochastic Recurrent Networks | Third Party | 2017 | 40+ |
| Latent Space Optimal Transport for Generative Models | No | 2018 | 10- |
| Improving explorability in variational inference with annealed variational objectives | No | 2018 | 10- |
| Taming VAEs | No | 2018 | 10+ |
| Semi-Amortized Variational Autoencoders | No | 2018 | 40+ |
| Avoiding Latent Variable Collapse with Generative Skip Models | No | 2018 | 15+ |
| Spherical Latent Spaces for Stable Variational Autoencoders | Yes | 2018 | 10+ |
| Unsupervised Discrete Sentence Representation Learning for Interpretable Neural Dialog Generation | No | 2018 | 10+ |
| Learning Latent Representations For Style Control And Transfer in End-To-End Speech Synthesis | Third Party | 2018 | 10- |
| The Mutual Autoencoder: Controlling Information in Latent Code Representations | No | 2018 | 10- |
| Hierarchicaly-Structured Variational Autoencoder For Long Text Generation | No | 2018 | 10- |
| Iterative Amortized Inference | No | 2018 | 20+ |
| BIVA: A very deep hierarchy of latent variables for generative modeling | No | 2019 | 10- |
| preventing posterior collapse with delta-VAEs | No | 2019 | 10- |
| Diagnosing and enhancing VAE model | No | 2019 | 10- |
| MAE: Mutual Posterior-Divergence Regularization For Variational Auto Encoder | No | 2019 | 10- |
| Topic-Guided Variational Autoencoders for Text Generation | No | 2019 | 10- |
| Importance Weighted Hierarchical Variational Inference | No | 2019 | 10- |
| Generated Loss, Augmented Training, And Multiscale VAE | No | 2019 | 40+ |
| mμ-Forcing: Training Variational Recurrent Autoencoders for Text Generation | No | 2019 | 10- |
| Dueling Decoders: Regularizing Variational Autoencoder Latent Spaces | No | 2019 | 10- |
| Riemannian Normalizing Flow on Variational Wasserstein Autoencoder for Text Modeling | No | 2019 | 10- |
| LIA: Latently Invertible Autoencoder with Adversarial Learning | Yes | 2019 | 10- |
| Compound Variational Auto-Encoder | No | 2019 | 10- |
| Quantization-Based Regularization for Autoencoders | No | 2019 | 10- |
| Lagging Ingerence Network And Posterior Collapse In Variational Autoencoders | Yes | 2019 | 10+ |
| Cyclical Annealing Schedule: A Simple Approach to Mitigating KL Vanishing | No | 2019 | 10- |
| Understanding Posterior Collapse in Generative Latent Variable Models | No | 2019 | 10- |
Some papers observed posterior collapse for a particular task and tried to alleviate it mostly by KL-annealing:
| Paper | Implementation | Publication Year | Citation |
|---|---|---|---|
| Generating Sentences from a Continuous Space | Third Party | 2015 | 600+ |
| A Neural Representation of Sketch Drawings | No | 2017 | 150+ |
| Improved Variational Autoencoders for Text Modeling using Dilated Convolutions | Third Party | 2017 | 80+ |
| Diverse and Accurate Image Description Using a Variational Auto-Encoder with an Additive Gaussian Encoding Space | Thrid Party | 2017 | 40+ |
| A Hierarchical Latent Vector Model for Learning Long-Term Structure in Music | Yes | 2018 | 50+ |
| Improving Variational Encoder-Decoders in Dialogue Generation | Third Party | 2018 | 50 |
| The challenge of realistic music generation: modelling raw audio at scale | No | 2018 | 20+ |
| Learning Product Codebooks Using Vector-Quantized Autoencoder For Image Retrieval | Yes | 2018 | 20+ |
| Auto-Encoding Variational Neural Machine Translation | No | 2018 | 20+ |
| Trajectory-User Linking via Variational AutoEncoder | No | 2018 | 20+ |
| Structure-aware Generative Network for 3D-Shape Modeling | No | 2018 | 20+ |
| Auto-Encoding Variational Neural Machine Translation | No | 2018 | 20+ |
| Unsupervised speech representation learning using WaveNet autoencoders | No | 2019 | 10- |
| Syntax-Infused Variational Autoencoder for Text Generation | No | 2019 | 10- |
| Unsupervised Recurrent Neural Network Grammars | No | 2019 | 10- |
| Learning Latent Plans from Play | No | 2019 | 10- |
| DialogWAE: Multimodal Response Generation With Conditional Wasserstein Auto-Encoder | No | 2019 | 10- |
Highly related papers but not exactly on Posterior Collapse:
| Paper | Implementation | Publication Year | Citation |
|---|---|---|---|
| Adversarial Autoencoders | Third Party | 2015 | 750+ |
| Improved variational inference with inverse autoregressive flow | Yes | 2016 | 400+ |
| Stick-Breaking VAE | No | 2016 | 40+ |
| Variational lossy autoencoder | No | 2016 | 190+ |
| Symmetrized Variational Inference | No | 2016 | 10- |
| ELBO surgery: yet another way to carve up the variational evidence lower bound | No | 2017 | 70+ |
| Adversarially Regularized Autoencoders | Yes | 2018 | 60+ |
| Adversarial Symmetric Variational Autoencoder | No | 2017 | 25+ |
| beta-VAE: Learning basic visual concepts with a constrained variational framework | Third Party | 2017 | 350+ |
| Distribution matching in variational inference | Third Party | 2017 | 15+ |
| Learning Latent Representations for Style Ccontrol And Transfer In end-to-end Speech Synthesis | Third Party | 2018 | 50+ |
| Wassertain Auto-Encoder: Latent Dimentionality And Random Encoders | No | 2018 | 10- |
| Learning Deep Representation by Mutual Information Estimation And Maximization | No | 2018 | 20+ |
| Sinkhorn AutoEncoders | No | 2018 | 10- |
| Learning Priors for Adversarial Autoencoders | No | 2018 | 10- |
| Hyperspherical Variational Auto-Encoders | Yes | 2018 | 30+ |
| Universal Audio Synthesizer Control With Normalizing Flows | Yes | 2018 | 10- |
| Representation Learning with Contrastive Predictive Coding | Third Party | 2018 | 50+ |