Google Research Introduces SensorFM: A Wearable Health Foundation Model Pretrained on One Trillion Minutes of Sensor Data
Google Research has developed SensorFM, a foundation model trained on one trillion minutes of sensor data to improve health tracking accuracy.
- Trained on a massive dataset of one trillion minutes of sensor signals.
- Uses a ViT-1D architecture to process time-series data from wearables.
- Outperformed traditional feature-engineered baselines on 34 out of 35 tasks.
- Demonstrates the effectiveness of foundation models in the health-tech domain.
SensorFM utilizes a ViT-1D masked-autoencoder backbone, allowing it to learn general representations from massive amounts of unlabeled sensor signals. The model was pretrained using data from 5 million consented participants, providing a scale of training data rarely seen in wearable health research.
Researchers tested the model across various sizes and data volumes to determine scaling laws. The results indicate that frozen embeddings combined with a simple linear probe outperformed traditional feature-engineered methods on nearly all 35 tested tasks.
To optimize performance, the team employed an agentic classroom approach to search through over 30,000 prediction heads. This allows the model to be fine-tuned for specific health predictions with high precision and efficiency.
Provides a blueprint for applying transformer-based foundation models to time-series sensor data.
Could lead to a new generation of highly accurate, generalized health-tracking products.
Signals Google's aggressive move into the intersection of AI and personalized healthcare.
Potential for more accurate early detection of health issues via consumer wearables.
- ViT-1D
- A one-dimensional Vision Transformer adapted for sequential sensor data instead of 2D images.
- Masked-autoencoder
- A self-supervised learning technique where the model learns by predicting missing parts of the input data.
Neuro-symbolic artificial intelligence in medicine - Nature
AI Cold War: Is Artificial Intelligence Driving the New Era of Great Power Competition? - Vision of Humanity
AI ResearchMeet LingBot-World-Infinity: An Open Causal World Model With An Agentic Harness
Artificial intelligence: University of Chicago Law School to ban phones, laptops in classroom for 1st-year students in AI strategy - ABC7 Chicago
San Andreas Fault: Hidden movements revealed by artificial intelligence - EurekAlert!
Accenture Federal Services Wins $821M CDAO Task Order for War Data Platform - GovCon Wire
Accenture Federal Services won an $821 million task order to build a data platform for the US Department of Defense's Chief Digital and Artificial Intelligence Office.
Comparison of YouTube Videos and Artificial Intelligence Chatbot Responses Regarding Pediatric Pelvic Floor Dysfunction: A Cross-Sectional Infodemiology Study - Cureus
A study in Cureus compared YouTube videos and AI chatbot responses on pediatric pelvic floor dysfunction, finding significant differences in accuracy and reliability.
Pope Leo XIV urges human-centred approach to artificial intelligence - The Catholic Herald
Pope Leo XIV has issued a call for artificial intelligence to be developed and used in ways that prioritize human dignity and ethical considerations.
BusinessHow Deutsche Telekom is rewiring telecommunications with AI
Deutsche Telekom is leveraging AI to transform its customer service, employee workflows, and network operations, partnering with OpenAI.
How Artificial Intelligence is Reshaping Real Estate - Bethpage Newsgram
Artificial intelligence is revolutionizing the real estate industry with new technologies and applications.
Robot Dogs, Teslas, and Rescue Helicopters: The UN AI Summit Was a Lot
The UN’s AI for Good summit highlighted the gap between rapid AI advancements and lagging global governance, featuring robotics, autonomous vehicles, and emergency response tech.