Isolating Sources of Disentanglement in Variational Autoencoders. Chen, R., T., Q., Li, X., Grosse, R., & Duvenaud, D. 2, 2018.
Isolating Sources of Disentanglement in Variational Autoencoders [link]Website  abstract   bibtex   2 downloads  
We decompose the evidence lower bound to show the existence of a term measuring the total correlation between latent variables. We use this to motivate our \$\textbackslashbeta\$-TCVAE (Total Correlation Variational Autoencoder), a refinement of the state-of-the-art \$\textbackslashbeta\$-VAE objective for learning disentangled representations, requiring no additional hyperparameters during training. We further propose a principled classifier-free measure of disentanglement called the mutual information gap (MIG). We perform extensive quantitative and qualitative experiments, in both restricted and non-restricted settings, and show a strong relation between total correlation and disentanglement, when the latent variables model is trained using our framework.
@article{
 title = {Isolating Sources of Disentanglement in Variational Autoencoders},
 type = {article},
 year = {2018},
 websites = {https://arxiv.org/abs/1802.04942v5},
 month = {2},
 id = {cadbac75-f2d0-3482-a13f-e5a01928323c},
 created = {2022-03-28T09:45:01.766Z},
 accessed = {2022-02-22},
 file_attached = {true},
 profile_id = {235249c2-3ed4-314a-b309-b1ea0330f5d9},
 group_id = {1ff583c0-be37-34fa-9c04-73c69437d354},
 last_modified = {2022-03-29T08:02:29.234Z},
 read = {false},
 starred = {false},
 authored = {false},
 confirmed = {true},
 hidden = {false},
 citation_key = {chenIsolatingSourcesDisentanglement2018},
 source_type = {article},
 private_publication = {false},
 abstract = {We decompose the evidence lower bound to show the existence of a term measuring the total correlation between latent variables. We use this to motivate our \$\textbackslashbeta\$-TCVAE (Total Correlation Variational Autoencoder), a refinement of the state-of-the-art \$\textbackslashbeta\$-VAE objective for learning disentangled representations, requiring no additional hyperparameters during training. We further propose a principled classifier-free measure of disentanglement called the mutual information gap (MIG). We perform extensive quantitative and qualitative experiments, in both restricted and non-restricted settings, and show a strong relation between total correlation and disentanglement, when the latent variables model is trained using our framework.},
 bibtype = {article},
 author = {Chen, Ricky T Q and Li, Xuechen and Grosse, Roger and Duvenaud, David}
}

Downloads: 2