VCR: Self-Supervised Framework for Incomplete Wearable Signals
A team of researchers has introduced VCR, a self-supervised learning framework designed to derive accurate representations resilient to absent modalities in data from wearable devices. Existing techniques frequently attempt to reconstruct missing signals, which can lead to fabricated details. VCR employs an orthogonal tokenizer to ensure clear separation through rectified latent manifolds and geometric projection, distinguishing each modality into common semantics and specific residuals. This method maintains the integrity of information while enhancing robustness. It effectively tackles the widespread issue of sensor incompleteness and the scarcity of labeled data in practical health monitoring scenarios.
Key facts
- VCR stands for Learning Valid Contextual Representation for Incomplete Wearable Signals.
- The paper is published on arXiv with ID 2605.18837.
- Wearable devices enable continuous health monitoring from multimodal signals.
- Real-world deployment is hindered by limited labeled data and pervasive sensor incompleteness.
- Most existing self-supervised methods assume full modality availability.
- Current approaches for handling missing modalities often reconstruct entire absent signals.
- Reconstruction can encourage hallucinating modality-specific details not inferable from observed signals.
- VCR employs an orthogonal tokenizer to enforce strict orthogonal disentanglement.
- The tokenizer rectifies latent manifolds and applies geometric projection.
- Each modality is separated into shared semantics and modality-specific residuals.
Entities
Institutions
- arXiv