[2601.06572] Hellinger Multimodal Variational Autoencoders

Computer Science > Machine Learning arXiv:2601.06572 (cs) [Submitted on 10 Jan 2026 (v1), last revised 30 Mar 2026 (this version, v2)] Title:Hellinger Multimodal Variational Autoencoders Authors:Huyen Vo, Isabel Valera View a PDF of the paper titled Hellinger Multimodal Variational Autoencoders, by Huyen Vo and 1 other authors View PDF HTML (experimental) Abstract:Multimodal variational autoencoders (VAEs) are widely used for weakly supervised generative learning with multiple modalities. Predominant methods aggregate unimodal inference distributions using either a product of experts (PoE), a mixture of experts (MoE), or their combinations to approximate the joint posterior. In this work, we revisit multimodal inference through the lens of probabilistic opinion pooling, an optimization-based approach. We start from Hölder pooling with $\alpha=0.5$, which corresponds to the unique symmetric member of the $\alpha\text{-divergence}$ family, and derive a moment-matching approximation, termed Hellinger. We then leverage such an approximation to propose HELVAE, a multimodal VAE that avoids sub-sampling, yielding an efficient yet effective model that: (i) learns more expressive latent representations as additional modalities are observed; and (ii) empirically achieves better trade-offs between generative coherence and quality, outperforming state-of-the-art multimodal VAE models. Comments: Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI) Cite as: arXiv:2601.065...