Multimodal Deep Generative Model for Semi-Supervised Learning Under Class Imbalance
A novel multimodal deep generative model tackles semi-supervised learning in the context of class imbalance. This method employs distinct encoders for each modality while allowing latent variables to be shared across them, and it streamlines joint posterior computation using a product-of-experts technique. To combat class imbalance more effectively, it substitutes standard Gaussian distributions with Student's t-distributions for the priors. The model aims to utilize complementary modalities and reduce bias transfer from imbalanced labeled data to pseudo-labels for the unlabeled dataset.
Key facts
- Proposes a multimodal deep generative model for semi-supervised learning under class imbalance
- Uses separate encoders for each modality with shared latent variables
- Employs product-of-experts method for joint posterior computation
- Replaces Gaussian distributions with Student's t-distributions for priors to address imbalance
- Aims to reduce bias propagation from imbalanced labeled data to pseudo-labels
- Published on arXiv with ID 2605.06289
- Announce type: cross
- Addresses the problem of class imbalance amplified under partial supervision
Entities
Institutions
- arXiv