On-device machine learning models play a key role in applications such as virtual assistants, document scanners, photo editors, facial recognition applications etc. on modern smartphones and other devices by eliminating the dependency on network connectivity and delivering a consistent and satisfactory user experience. However, the embedded model architecture makes the task of flighting a new machine learning model challenging. In model flighting, a small portion of the user base is exposed to the new version of the model, while the rest of the user base continues to use the old version of the model that is already part of the application. The performance of both models is compared and the model that has better performance is chosen for deployment in the main application. This disclosure describes an approach to seamlessly conduct flighting for machine learning models embedded in smart devices on hundreds of thousands smart devices instead of a much smaller sample of in-house devices in a lab environment. Using the architecture described in this disclosure, model flighting can be achieved without a user downloading and updating to a new version of the application.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.