How AI Models Learn from Your Health Data Without Sharing It

This patent describes a system where personal health data, like photos or audio, stays on your device while helping train a shared AI model for medical research, improving predictions without sending your raw data.

What it covers

This system uses a method called federated learning to train artificial intelligence models. It involves multiple 'edge devices' (like your smartphone) that each have a program called a 'federated learner' running a 'deep neural network' (DNN). This DNN predicts characteristics about you, such as your mood or age, from your recorded data like pictures or audio, as mentioned in claims 1, 2, 5, and 7. If the prediction is wrong, you can correct it. Your device then uses your private data and corrections to update its part of the DNN model. Instead of sending your raw data, only a 'modified part of the modified tensor' (a small, updated piece of the AI model) is sent to a central 'FL aggregator'. The aggregator combines these updates from many devices to create a better, shared model, which it then sends back to all the devices. For example, claim 1 specifies that if the data is a picture, the system only performs this update training if a face is detected in the image.

What it doesn't cover

• —It does not cover systems where raw user data, like pictures or audio recordings, is directly uploaded to a central server for training.
• —It does not cover AI training systems that do not perform 'update training' of the deep neural network directly on the end user's device.
• —It does not cover federated learning systems that send the entire model parameters from the edge device to the aggregator, rather than just a 'modified part'.
• —It does not cover systems that use picture data for training without first detecting a face in the picture, as specified in claim 1.
• —It does not cover federated learning without an 'FL aggregator' that combines updates from multiple edge devices to produce a new base model.

The clever bit

The clever bit is enabling AI models to learn from highly sensitive personal data, like facial images or voice recordings, directly on a user's device. This on-device training, combined with the selective sharing of only 'modified parts' of the model, allows for continuous improvement of a shared AI without ever exposing individual raw data to a central server. The specific requirement in claim 1 to only train on pictures if a face is detected adds a layer of focused privacy and relevance.

Why it matters

This patent addresses a significant challenge in medical research: how to use vast amounts of sensitive personal health data to train powerful AI models without compromising individual privacy. By keeping raw data on user devices and only sharing model updates, it enables the development of more accurate diagnostic and predictive tools. This approach is crucial for advancing personalized medicine and public health initiatives while adhering to strict data privacy regulations.

Real-world examples

• 1.Google's Gboard keyboard for next-word prediction
• 2.Apple's Siri voice recognition improvements
• 3.NVIDIA Clara Federated Learning platform for healthcare
• 4.Various research initiatives in personalized health and diagnostics
• 5.Sharecare AI applications for health management