Research map / Vision-language and multimodal / Cross-modal learning
Learning Multimodal VAEs through Mutual Supervision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 mutual supervision instead of MVAE | Multimodal VAE · MNIST-SVHN, MNIST→SVHN, f = 1.0 | cross coherence: 0.123 ± 0.003 -> 0.625 ± 0.007 | Helped |
| Use mutual supervision instead of MVAE | Multimodal VAE · MNIST-SVHN, SVHN→MNIST, f = 1.0 | cross coherence: 0.498 ± 0.100 -> 0.752 ± 0.004 | Helped |
| Use mutual supervision instead of MVAE | Multimodal VAE · MNIST-SVHN, MNIST latent classification, f = 1.0 | accuracy: 0.892 ± 0.005 -> 0.908 ± 0.007 | Helped |
| Use mutual supervision instead of MVAE | Multimodal VAE · MNIST-SVHN, SVHN latent classification, f = 1.0 | accuracy: 0.131 ± 0.010 -> 0.648 ± 0.012 | Helped |
| Use mutual supervision instead of MMVAE | Multimodal VAE · CUB, image→captions, f = 1.0 | CCA correlation: 0.060 ± 0.010 -> 0.064 ± 0.011 | Helped |
| Use mutual supervision instead of MMVAE | Multimodal VAE · CUB, captions→image, f = 1.0 | CCA correlation: 0.058 ± 0.001 -> 0.074 ± 0.001 | Helped |
| Reduce the number of pseudo-samples used to estimate the prior (ablation) | MEME · Partially observed multimodal data | performance | Hurt |
| Add partially observed training data alongside fully observed data (ablation) | MEME · Partially observed multimodal data | performance | Helped |
| Use the bidirectional objective with shared parameters instead of training in one direction (ablation) | MEME · MNIST-SVHN | marginal log likelihood: -428728.7 -> -11355.4 | Helped |
| Stop gradients through the probability ratio in the encoder estimator (ablation) | MEME · Encoder gradient estimation | signal-to-noise ratio | Helped |
| Include the approximate-posterior sample in the LogSumExp estimate (ablation) | MEME · Numerical computation of the probability ratio | numerical stability | Helped |
Closest papers by meaning
- Unity by Diversity: Improved Representation Learning for Multimodal VAEs (NeurIPS 2024 · similarity 0.90)
- Disentangled Cross-Modal Representation Learning with Enhanced Mutual Supervision (NeurIPS 2025 · similarity 0.89)
- On the Limitations of Multimodal VAEs (ICLR 2022 · similarity 0.89)
- Multimodal Learning Without Labeled Multimodal Data: Guarantees and Applications (ICLR 2024 · similarity 0.87)
- InfMasking: Unleashing Synergistic Information by Contrastive Multimodal Interactions (NeurIPS 2025 · similarity 0.86)
- What to align in multimodal contrastive learning? (ICLR 2025 · similarity 0.86)
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.