Learning Disentangled Joint Continuous and Discrete Representations
arXiv:1804.00104
Abstract
We present a framework for learning disentangled and interpretable jointly continuous and discrete representations in an unsupervised manner. By augmenting the continuous latent distribution of variational autoencoders with a relaxed discrete distribution and controlling the amount of information encoded in each latent unit, we show how continuous and categorical factors of variation can be discovered automatically from data. Experiments show that the framework disentangles continuous and discrete generative factors on various datasets and outperforms current disentangling methods when a discrete generative factor is prominent.
NIPS camera ready, added quantitative evaluation and figures for dsprites dataset
References in corpus (6)
- Fashion-MNIST: a Novel Image Dataset for Benchmarking Machine Learning Algorithms
- The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables
- Deep Convolutional Inverse Graphics Network
- Disentangling by Factorising
- Isolating Sources of Disentanglement in Variational Autoencoders
- Auto-Encoding Total Correlation Explanation
Cited by in corpus (17)
- Recent Advances in Autoencoder-Based Representation Learning
- Disentangling Hate in Online Memes
- Relevance Factor VAE: Learning and Identifying Disentangled Factors
- Dual Swap Disentangling
- InfoCatVAE: Representation Learning with Categorical Variational Autoencoders
- Disentangling Controllable and Uncontrollable Factors of Variation by Interacting with the World
- Learning latent representations across multiple data domains using Lifelong VAEGAN
- Disentangling and Learning Robust Representations with Natural Clustering
- A Biologically Inspired Visual Working Memory for Deep Networks
- Double cycle-consistent generative adversarial network for unsupervised conditional generation
- Unsupervised Learning of Neurosymbolic Encoders
- Physical discovery in representation learning via conditioning on prior knowledge: applications for ferroelectric domain dynamics
- TzK: Flow-Based Conditional Generative Model
- Disassembling Object Representations without Labels
- Differentiable Disentanglement Filter: an Application Agnostic Core Concept Discovery Probe
- Polyline Generative Navigable Space Segmentation for Autonomous Visual Navigation
- DEFT: Distilling Entangled Factors by Preventing Information Diffusion