Variational Autoencoders with Decremental Information Bottleneck for Disentanglement
Jiantao Wu (University of Surrey),* Shentong Mo (Carnegie Mellon University), Xingshen Zhang (University of Jinan), Muhammad Awais (University of Surrey), Sara Ahmed (University of surrey), Zhenhua Feng (University of Surrey), Lin Wang (University of Jinan), Xiang Yang (Zhejiang Mingyi Technology Co., Ltd.)
The 34th British Machine Vision Conference
Abstract
One major challenge of disentanglement learning with variational autoencoders is the trade-off between disentanglement and reconstruction fidelity. Previous studies, which increase the information bottleneck during training, tend to lose the constraint of disentanglement, leading to the information diffusion problem. In this paper, we present a novel framework for disentangled representation learning, DeVAE, which utilizes hierarchical latent spaces with decreasing information bottlenecks across these spaces. The key innovation of our approach lies in connecting the hierarchical latent spaces through disentanglement-invariant transformations, allowing the sharing of disentanglement properties among spaces while maintaining an acceptable level of reconstruction performance. Through a series of experiments and ablation studies on dSprites and Shapes3D, we demonstrate the effectiveness of DeVAE in achieving a balance between disentanglement and reconstruction.Video
Citation
@inproceedings{Wu_2023_BMVC,
author = {Jiantao Wu and Shentong Mo and Xingshen Zhang and Muhammad Awais and Sara Ahmed and Zhenhua Feng and Lin Wang and Xiang Yang},
title = {Variational Autoencoders with Decremental Information Bottleneck for Disentanglement},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0433.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by