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.


    Access

    Download


    Export, share and cite



    Title :

    Mitigating Modality Collapse in Multimodal VAEs via Impartial Optimization


    Contributors:

    Publication date :

    2022


    Remarks:

    Accepted as a Spotlight paper at ICML 2022. 27 pages, 10 figures


    Type of media :

    Preprint


    Type of material :

    Electronic Resource


    Language :

    English




    Mitigating Modality Collapse in Multimodal VAEs Via Impartial Optimization

    Javaloy, Adrian / Meghdadi, Maryam / Valera, Isabel | TIBKAT | 2023


    On the Limitations of Multimodal VAEs

    Daunhawer, Imant / Sutter, Thomas M. / Chin-Cheong, Kieran et al. | ArXiv | 2021

    Free access

    f-VAEs: Improve VAEs with Conditional Flows

    Su, Jianlin / Wu, Guang | ArXiv | 2018

    Free access

    On Posterior Collapse and Encoder Feature Dispersion in Sequence VAEs

    Long, Teng / Cao, Yanshuai / Cheung, Jackie Chi Kit | ArXiv | 2019

    Free access

    LEARNING HIERARCHICAL PRIORS IN VAES

    Klushyn, Alexej / Chen, Nutan / Kurle, Richard et al. | TIBKAT | 2020