NASA’s new dark energy space telescope can also detect killer asteroids
technologyreview.com·3h ago
AI Research
arXiv CS.AI·May 21·6 min read
VCR: Learning Valid Contextual Representation for Incomplete Wearable Signals
TL;DR
Wearable health sensors produce multimodal data (signals from multiple sensor types), but real-world deployment fails when sensors malfunction or data is missing and labeled training examples are scarce. VCR (Valid Contextual Representation) uses self-supervised pretraining to learn from unlabeled wearable data while handling incomplete modalities without reconstructing absent signals, which prevents the model from inventing fake sensor readings.
✦ Why It Matters
Engineers deploying wearable health systems can now handle real-world sensor failures without expensive labeling or hallucinated data artifacts.
Key Takeaways
- Avoid reconstructing missing sensor data; learn directly from incomplete multimodal signals instead.
- Use contrastive self-supervised learning to extract patterns from available modalities only.
- VCR handles variable sensor availability without retraining for each failure scenario.
- Test robustness across different missingness patterns before deploying to production wearables.
Full Summary
Wearable devices continuously monitor health through multiple sensor types (accelerometers, heart rate monitors, etc.), but real-world systems suffer from incomplete modality availability—sensors fail, are removed, or unavailable. Self-supervised pretraining reduces dependence on labeled data, but existing methods assume all modalities are present.
Prior approaches reconstruct missing signals entirely, which can cause models to learn spurious patterns from fabricated data rather than genuine sensor relationships. VCR learns valid contextual representations by training on incomplete signals directly, without reconstruction.
The method uses contrastive learning (comparing similar and dissimilar examples) to extract meaningful patterns from available modalities only. Experiments demonstrate VCR achieves better downstream task performance than reconstruction-based methods while maintaining robustness to varying levels of sensor missingness.