Abstract
A number of variational autoencoders (VAEs) have recently emerged with the aim of modeling multimodal data, e.g., to jointly model images and their corresponding captions. Still, multimodal VAEs tend to focus solely on a subset of the modalities, e.g., by fitting the image while neglecting the caption. We refer to this limitation as modality collapse. In this work, we argue that this effect is a consequence of conflicting gradients during multimodal VAE training. We show how to detect the sub-graphs in the computational graphs where gradients conflict (impartiality blocks), as well as how to leverage existing gradient-conflict solutions from multitask learning to mitigate modality collapse. That is, to ensure impartial optimization across modalities. We apply our training framework to several multimodal VAE models, losses and datasets from the literature, and empirically show that our framework significantly improves the reconstruction performance, conditional generation, and coherence of the latent space across modalities.
Original language | English |
---|---|
Title of host publication | Proceedings of the 39th International Conference on Machine Learning |
Editors | Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, Sivan Sabato |
Publisher | PMLR |
Pages | 9938-9964 |
Number of pages | 27 |
Volume | 162 |
Publication status | Published - 23 Jul 2022 |
Event | 39th International Conference on Machine Learning - Baltimore, United States Duration: 17 Jul 2022 → 23 Jul 2022 Conference number: 39 https://icml.cc/Conferences/2022 |
Conference
Conference | 39th International Conference on Machine Learning |
---|---|
Abbreviated title | ICML 2022 |
Country/Territory | United States |
City | Baltimore |
Period | 17/07/22 → 23/07/22 |
Internet address |