Research map / Vision-language and multimodal / Human and embodied modeling

Scaling Wearable Foundation Models

ICLR 2025 · Vision-language and multimodal › Human and embodied modeling · approach: Medical vision-language models · 5 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a masked-autoencoder wearable sensor foundation model (LSM) on up to 40M hours of multimodal Fitbit/Pixel Watch data from 165K users and studies its scaling with compute, data, and model size on imputation, forecasting, and classification tasks.

Problem

Wearable sensor data is unlabeled, imbalanced, and collected from small populations; it is unclear whether neural scaling laws hold for multimodal wearable sensor data and how well such foundation models transfer to generative and discriminative tasks.

Approach

Applies masked signal modeling (MAE-style, 80% random patch masking over a 26-signal x 300-minute image) to large-scale wearable data, systematically measures power-law scaling with compute, data, and model size, and compares against contrastive SSL baselines (MSN, DINO, SimCLR).

Key results

• LSM improves temporal interpolation MAE by 16% to 23% and extrapolation MAE by 20% to 21% over the best baseline across 10 to 120 minute horizons.
• Fine-tuned LSM reaches 90.3% accuracy and 97.0 mAP on exercise detection, a 27% / 57% gain over supervised training.
• Fine-tuned LSM reaches 68.5% accuracy and 51.4 mAP on 8-class activity recognition, a 29% / 54% gain over supervised training.

Measured effects reported in the paper

Surprising findings

• Scaling beyond about 10^7 hours of data and roughly 10^8 parameters yields saturating returns; the 40M-hour model did not improve over 6.6M hours.
• Subject count and hours per subject can be traded off, with total hours driving gains more than the number of subjects.

Limitations

• Scaling saturates at about 10^7 hours and 10^8 parameters, which the authors attribute to the masked pretext task and dataset construction.
• Heavy reliance on imputation for missing wearable data, with data likely missing not at random.
• Limited discriminative evaluation: only exercise detection and activity recognition, which the authors acknowledge.

Experimental setup

Datasets: Fitbit Sense 2 / Pixel Watch 2 sensor data (165,090 subjects, Jan 2023 to Jul 2024) user-labeled exercise and activity downstream sets

Benchmarks: temporal interpolation temporal extrapolation random imputation sensor imputation exercise detection activity recognition (8-class)

Metrics: MAE MSE accuracy mAP

Baselines: Mean Nearest Neighbor Linear Interpolation MSN DINO SimCLR Supervised ViT

Builds on: masked autoencoder MSN DINO SimCLR T5 model size conventions

Closest papers by meaning

• Beyond Sensor Data: Foundation Models of Behavioral Data from Wearables Improve Health Predictions (ICML 2025 · similarity 0.91)
• Inertia-1: An Open Exploration of Wearable Motion Foundation Models (NeurIPS 2026 · similarity 0.89)
• OpenMHC: Accelerating the Science of Wearable Foundation Models (NeurIPS 2026 · similarity 0.89)
• RelCon: Relative Contrastive Learning for a Motion Foundation Model for Wearable Data (ICLR 2025 · similarity 0.88)
• Large-scale Training of Foundation Models for Wearable Biosignals (ICLR 2024 · similarity 0.87)
• SensorLM: Learning the Language of Wearable Sensors (NeurIPS 2025 · similarity 0.86)