VAE
VAE 最通常联系的topic有:
- Disentangled Representation Learning.
- Image Generation. 趋势是VAE和GAN模型的融合,优势互补。
unsorted
- AutoEncoding Variational Bayes. [paper] (Origninal Paper 1).
- Stochastic backpropagation and approximate inference in deep generative models. [paper] (Origninal Paper 2)
- Tutorial on Variational Autoencoders. [paper] CMU和UCB的VAE的学习材料,阐述了基本的VAE思想,还包括CVAE的内容。
- Learning Structured Output Representation using Deep Conditional Generative Models. [paper] (CVAE)。
- Adversarial AutoEncoder. [paper] (AAE)
- Wasserstein Auto-Encoders. [paper]
- Adversarial Latent Autoencoders. [paper] (ALAE) 近来火热的SOTA人脸生成模型。
- Importance Weighted AutoEncoders. [paper]. 基于重要性采样的 VAE 的改进。
Disentangled Representation Learning
Sparsity Problem
Sparsity Problem是VAE类模型的常见问题:VAE的编码器仅使用隐变量的小子集。这在《Deep Learning》一书中也被提及到。该问题也被理解为VAE模型会自剪枝(self-pruning)、过剪枝(over-pruning)、后验坍缩(posterior collapse)问题。
和隐变量探索(latent variable collapse)区分开:隐变量探索指的是当近似后验完全和先验高斯相等时,近似后验完全和输入无关,什么也学不到。
这或许是VAE类模型用于多模态数据建模的正常现象?
- Tackling Over-pruning in Variational Autoencoders. [paper] !!!
- Sparsity in Variational Autoencoders。 [paper] !!!
- Variational Autoencoders and the Variable
Collapse Phenomenon. [paper] 这和上面是同一篇文章。 - Redundancy-resistant Generative Hashing for Image Retrieval. [paper] !!
Likely
- Variance Loss in Variational Autoencoders. [paper]
- Don’t Blame the ELBO! A Linear VAE Perspective on Posterior Collapse. [paper] !!
- Improved Variational Inference with Inverse Autoregressive Flow. [paper] 同时本文也是Kingma本人的一作。 !
- Deep Unsupervised Clustering with Gaussian Mixture Variational Autoencoders. [paper] 这篇文章或许关注了该问题的出现是因为VAE对多模态数据建模? !!
- Avoiding Latent Variable Collapse With Generative Skip Models. [paper] 通过加入skip-connection的结构缓解问题。
- Lagging Inference Networks and Posterior Collapse in Variational Autoencoders. [paper] 认为后验坍缩是因为在训练的早期阶段,encoder推断的近似后验落后于真实后导验致的。
- Generating sentences from a continuous space. [paper] 其中3.1小节涉及到解决后验坍缩问题。
- How to train deep variational autoencoders and probabilistic ladder networks. [paper] 在Dissucsion的
Latent Representation中涉及到sparse问题。