OmniDecVAEs: Full-Stack Wearable Disentangled Representations
Researchers have introduced a new system called Omni-modal Variational Decomposition Autoencoders, or OmniDecVAEs, designed to pull apart and understand data from various wearable sensors. This method addresses the limitations of existing approaches by combining classification, clear representation learning, data fusion, and generative modeling for different time series. Building on earlier DecVAEs, OmniDecVAEs use a multi-view self-supervised decomposition loss alongside a unique asymmetric autoencoder setup to learn time-frequency subspaces tailored to each modality. This adaptable framework can accommodate multiple data types. Results from a challenging human activity recognition task demonstrate its effectiveness. You can find the complete research paper on arXiv with the ID 2608.07385.
Key facts
- OmniDecVAEs is a new framework for learning disentangled representations in wearable computing.
- It addresses task-specific classification, interpretable representation learning, fusion, and generative modeling simultaneously.
- The framework extends DecVAEs with modality-conditioned time-frequency latent subspaces.
- It uses a multi-view self-supervised decomposition loss and a shared asymmetric autoencoder.
- The method is designed for arbitrarily many modalities and is scalable and unified.
- Results are reported on a challenging omni-modal human activity recognition task.
- The paper is available on arXiv with identifier 2608.07385.
Entities
Institutions
- arXiv