講演抄録/キーワード |
講演名 |
2020-01-24 10:10
Optimal Transport based Autoencoder for class and style Disentanglement ○Florian Tambon・Tetsuo Furukawa(Kyutech) NC2019-62 |
抄録 |
(和) |
The Sinkhorn autoencoder is a novel generative model using optimal transport to model the aggregated posterior from samples, hence discarding traditional reparametrization trick from classical Variational Autoencoder (VAE) and allowing better flexibility of metrics spaces and priors. Yet, one of the down side of all latent space modelling methods is the lack of interpretability and the potential entanglement problem. The aim of this work is to extend the Sinkhorn Autoencoder to better disentangle the latent space by focusing on the class/style separation approach while providing better interpretability and generative capability. Thus, our method would help further expand knowledge regarding optimal transport based generative model. |
(英) |
The Sinkhorn autoencoder is a novel generative model using optimal transport to model the aggregated posterior from samples, hence discarding traditional reparametrization trick from classical Variational Autoencoder (VAE) and allowing better flexibility of metrics spaces and priors. Yet, one of the down side of all latent space modelling methods is the lack of interpretability and the potential entanglement problem. The aim of this work is to extend the Sinkhorn Autoencoder to better disentangle the latent space by focusing on the class/style separation approach while providing better interpretability and generative capability. Thus, our method would help further expand knowledge regarding optimal transport based generative model. |
キーワード |
(和) |
Generative model / Optimal Transport / Disentanglement / Sinkhorn Loss / Autoencoder / Latent Space / / |
(英) |
Generative model / Optimal Transport / Disentanglement / Sinkhorn Loss / Autoencoder / Latent Space / / |
文献情報 |
信学技報, vol. 119, no. 382, NC2019-62, pp. 17-22, 2020年1月. |
資料番号 |
NC2019-62 |
発行日 |
2020-01-16 (NC) |
ISSN |
Online edition: ISSN 2432-6380 |
著作権に ついて |
技術研究報告に掲載された論文の著作権は電子情報通信学会に帰属します.(許諾番号:10GA0019/12GB0052/13GB0056/17GB0034/18GB0034) |
PDFダウンロード |
NC2019-62 |