Google brings ML to Play services to the device on Android

Machine learning on the device offers a number of benefits, from speed (lower latency) to privacy. To help increase Android adoption, Google is now adding an “Android ML platform” – primarily TensorFlow Lite – directly to Play Services.

Google Play services are responsible for key features for users on Android and provide third-party application developers with access to various tools. The latest will be machine learning on the device.

The company has found several constraints that prevent applications from moving away from ML that occurs in the cloud. This includes the file size constraints of bundling additional libraries for ML at varying performance across devices, resulting in significant differences in stability and accuracy. Finally, “optimizing the scope can lead to the use of older and more widely available APIs”.

Google’s solution to this is a “fully integrated and updatable ML inference stack” called the “Android ML Platform”. It has three components, starting with Google which makes inference capabilities directly available on the device to almost all Android devices:

TensorFlow Lite will be available on all devices with Google Play services. Developers will no longer need to include the runtime in their applications, thereby reducing the size of the application.

To ensure “peak” performance on all devices, a new auto-acceleration feature in TensorFlowLite for Android enables “model testing to create authorization lists for specific devices taking into account performance, accuracy and stability ”. Available later this year, this determines whether hardware acceleration is enabled or not.

Finally, Google will update the Neural Networks API outside of Android operating system versions. It also works with chipset vendors, like Qualcomm, to deliver the latest device drivers aside from operating system updates.

This will allow developers to drastically reduce testing of thousands of devices to a handful of configurations. We are excited to announce that we will be launching later this year with Qualcomm as our first partner.