dpVAEs: Fixing Sample Generation for Regularized VAEs.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Unsupervised representation learning is crucial for computer vision tasks lacking labeled data.
• Variational Autoencoders (VAEs) are generative models for learning representations, but often struggle with downstream tasks.
• Regularization techniques improve VAE representations but degrade sample generation quality.

Purpose of the Study:

• To address the trade-off between representation learning and sample generation in regularized VAEs.
• To introduce a novel approach, decoupled priors (dpVAEs), to improve VAE representation learning.
• To enable effective use of VAE regularizers without compromising generative capabilities.

Main Methods:

• Introduced decoupled priors (dpVAEs) to separate representation and generation spaces.
• Utilized invertible networks for bijective mapping from complex to simple distributions.
• Adapted dpVAEs with existing state-of-the-art VAE regularizers.

Main Results:

• dpVAEs successfully decouple representation and generation spaces.
• Representation learning benefits of VAE regularizers are achieved without sacrificing sample generation.
• Experiments on MNIST, SVHN, and CelebA datasets demonstrate significant improvements.

Conclusions:

• dpVAEs resolve the representation-generation trade-off in regularized VAEs.
• This novel approach enhances unsupervised representation learning capabilities.
• dpVAEs offer a flexible and effective solution for improving VAE performance across various datasets.