Variational autoencoders

Mitigating Modality Collapse in Multimodal VAEs via Impartial Optimization
We connect the problem of modality collapse with conflicting gradients, and leverage solutions from multitask learning to alleviate it.
Adrián Javaloy, Maryam Meghdadi, Isabel Valera
Cite Code Poster Slides Video arXiv thread Proceedings
Mitigating Modality Collapse in Multimodal VAEs via Impartial Optimization
Boosting heterogeneous VAEs via multi-objective optimization
By leveraging MTL solutions for negative transfer, we improve the training dynamics of variational autoencoders (VAEs), boosting their performance on heterogeneous datasets.
Adrián Javaloy, Maryam Meghdadi, Isabel Valera
PDF Poster Slides
Boosting heterogeneous VAEs via multi-objective optimization