Thanks @thx.bsky.social , exciting times!

Still waiting for Google to open source the code! In the meantime, here is the paper: arxiv.org/pdf/2410.13638

Conclusion: Google’s LSM demonstrates the potential of foundation models in wearable tech, offering a more scalable, robust, and efficient way to interpret complex sensor data. This is a significant step forward in health #LSM #PaveTheWay #Health4All

Implications: Scaling multimodal models like LSM paves the way for more generalizable health monitoring tools. It could improve fitness tracking, disease detection, and personalization, while minimizing the need for extensive labeled datasets. #FoundationModels

Generative Power:
Beyond imputation, LSM excels in:
Temporal extrapolation: Predicting future sensor trends.
Sensor imputation: Filling gaps across multiple sensor modalities. These capabilities will enhance reliability and adaptability for health applications. #MaskAndPredict

Few-shot Learning: LSM shows significant label efficiency. For activity recognition (e.g., biking, walking), it outperforms supervised baselines with just 5-10 labeled examples per class, making it practical for diverse downstream applications. #fewshotslearning #activityrecognition

Scaling Law: as data, compute, and parameters increase, performance improves sublinearly. Gains saturate beyond ~10^7 data hours, consistent with NLP and vision models. Larger models like ViT-110M benefit most, requiring vast datasets to avoid overfitting #ScalingLaw

Imputation Improvements:
LSM reduces errors by up to 38% (MAE) over traditional methods like linear interpolation.
Missing data, caused by battery-saving modes or device fit, is common in wearables. LSM ensures robustness despite incomplete streams.

Core Tasks:
- Imputation: Reconstruct missing data due to device interruptions.
- Interpolation: Fill gaps between existing observations.
- Extrapolation: Predict future values for sensor signals.

The Dataset: LSM was trained on data from 165k Fitbit Sense 2 and Google Pixel Watch 2 users. It combines 40M+ hours of electrodermal activity (EDA), accelerometer, photoplethysmography (PPG), and skin temperature signals—the largest wearable dataset to date. #Dataset #Wearable