English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Multimodal variational autoencoders (VAEs) have shown promise as efficientgenerative models for weakly-supervised data. Yet, despite their advantage ofweak supervision, they exhibit a gap in generative quality compared to unimodalVAEs, which are completely unsupervised. In an attempt to explain this gap, weuncover a fundamental limitation that applies to a large family ofmixture-based multimodal VAEs. We prove that the sub-sampling of modalitiesenforces an undesirable upper bound on the multimodal ELBO and thereby limitsthe generative quality of the respective models. Empirically, we showcase thegenerative quality gap on both synthetic and real data and present thetradeoffs between different variants of multimodal VAEs. We find that none ofthe existing approaches fulfills all desired criteria of an effectivemultimodal generative model when applied on more complex datasets than thoseused in previous benchmarks. In summary, we identify, formalize, and validatefundamental limitations of VAE-based approaches for modeling weakly-superviseddata and discuss implications for real-world applications.

On the Limitations of Multimodal VAEs