Research map / Probabilistic and scientific modeling / Variational and generative sampling
Multimodal Generative Learning Utilizing Jensen-Shannon-Divergence
Thomas Sutter, Imant Daunhawer, Julia Vogt
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the mmJSD objective instead of MMVAE's objective | multimodal VAE · MNIST-SVHN-Text | latent representation accuracy (Joint): 0.93 -> 0.98 | Improved |
| Use the mmJSD objective instead of MVAE's objective | multimodal VAE · MNIST-SVHN-Text | generated-sample classification accuracy (M from S,T): 0.22 -> 0.95 | Improved |
| Use the mmJSD objective instead of MMVAE's objective | multimodal VAE · MNIST-SVHN-Text | generated-sample classification accuracy (M from S,T): 0.91 -> 0.95 | Improved |
| Add modality-specific latent subspaces (ablation) | mmJSD · MNIST-SVHN-Text | generated-sample precision (M, S,T condition): 0.16 -> 0.64 | Improved |
| Add modality-specific latent subspaces (ablation) | mmJSD · MNIST-SVHN-Text | latent representation accuracy (Joint): 0.98 -> 0.99 | Improved |
| Add modality-specific latent subspaces (ablation) | multimodal VAE · MNIST-SVHN-Text | generated-sample precision (M, S,T condition): mmJSD: 0.16 -> 0.64; MMVAE: 0.18 -> 0.63 | Mixed |
| Use mmJSD instead of MMVAE | multimodal VAE · CelebA | latent representation mean average precision (Joint): 0.42 -> 0.57 | Improved |
| Use mmJSD instead of MVAE | multimodal VAE · CelebA | generation mean average precision (T to I): 0.30 -> 0.42 | Improved |
| Use mmJSD instead of MMVAE | multimodal VAE · CelebA | generation mean average precision (T to I): 0.36 -> 0.42 | Improved |
| Use mmJSD instead of MVAE | multimodal VAE · MNIST-SVHN-Text | test set log-likelihood (joint model): -1864 -> -1961 | Hurt |
| Use modality-specific latent subspaces (ablation) | MMVAE · MNIST-SVHN-Text | generated-sample classification accuracy (S from M): 0.32 -> 0.88 | Improved |
| Use modality-specific latent subspaces (ablation) | mmJSD · MNIST-SVHN-Text | generated-sample classification accuracy (S from M): 0.37 -> 0.89 | Improved |
Closest papers by meaning
- Generalized Multimodal ELBO (ICLR 2021 · similarity 0.89)
- Multimodal Representation Learning by Alternating Unimodal Adaptation (CVPR 2024 · similarity 0.85)
- Aggregation of Dependent Expert Distributions in Multimodal Variational Autoencoders (ICML 2025 · similarity 0.85)
- Classifier-guided Gradient Modulation for Enhanced Multimodal Learning (NeurIPS 2024 · similarity 0.84)
- MMPareto: Boosting Multimodal Learning with Innocent Unimodal Assistance (ICML 2024 · similarity 0.84)
- Improving Multimodal Learning Balance and Sufficiency through Data Remixing (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
