The usage of machine learning (ML) has become quite common across various applications for multiple use-cases. While on-device ML is preferred over its server-based counterpart for a number of reasons such as low latency and lack of dependency on internet connectivity, it also has numerous drawbacks. Google says that it will be addressing these challenges with the Android ML Platform, coming later this year.
Some of the problems that developer face when implementing on-device machine learning solutions in their apps is that they increase the size of the APK, performance can take a hit on less powerful hardware, and there is a lack of more advanced features because developers tend to utilize older APIs to target more devices and users.
Google will tackle all these issues with its Android ML Platform, which it is boasting as an updateable and integrated on-device inference stack. APKs will not have to include machine learning libraries because the binaries for TensorFlow Lite will already be present in the OS. Performance will be managed and optimized by the system, which includes automatic triggering of hardware acceleration, among other things. Moreover, it will feature a standard Neural Networks API that spans across multiple Android versions.
More importantly, this API will be decoupled from the OS in the sense that updates can be offered via Google Play Services instead of having to release an OS update. In the same vein, Google is also collaborating with chipset manufacturers so that driver updates can be delivered in the same way and the dependency on OS updates is reduced.
Google says that the Android ML Platform will be made available later this year, but those interested in testing it sooner can sign up for the early access program here.