Construct smarter Android apps with on-device Machine Studying

Posted by Thomas Ezan, Developer Relations

Up to now 12 months, the Android group made vital enhancements to on-device machine studying to assist builders create smarter apps with extra options to course of photos, sound, and textual content. Within the Google I/O speak Construct smarter Android apps with on-device Machine Studying, David Miro-Llopis PM on ML Package and Thomas Ezan Android Developer Relation Engineer evaluate new Android APIs and options and showcase purposes utilizing on-device ML.

Working ML processes on-device permits low-latency, will increase data-privacy, permits offline help and doubtlessly reduces cloud invoice. Functions akin to Lens AR Translate or the doc scanning function obtainable in Information in India, profit from the benefits of on-device ML.

To deploy ML options on Android, builders have two choices:

ML Package: which presents production-ready ML options to widespread consumer flows, by way of easy-to-use APIs.

Android’s customized ML stack: which is constructed on high of Tensorflow Lite, and gives management over the inference course of and the consumer expertise.

ML Package launched new APIs and improved present options

During the last 12 months, the ML Package group labored on each bettering present APIs and launching new ones: face mesh and doc scanner. ML Package is launching a brand new doc scanner API in Q3 2023, that may present a constant scanning expertise throughout apps in Android. Builders will be capable of use it solely with a number of traces of code, while not having digital camera permission and with low apk measurement impression (on condition that will probably be distributed by way of Google Play Companies. Similarly, Google code scanner is now typically obtainable and gives a constant scanning expertise throughout apps, while not having digital camera permission, by way of Google Play Companies.

Moreover, ML Package improved the efficiency of the next APIs: barcode detection (by 17%), textual content recognition, digital ink recognition, pose detection, translation, and good reply. ML Package additionally built-in some APIs to Google Play Companies so that you don’t must bundle the fashions to your software. Many builders are utilizing ML Package to simply combine machine studying into their apps; for instance, WPS makes use of ML Package to translate textual content in 43 languages and save $65M a 12 months.

Acceleration Service in Android’s customized ML stack is now in public beta

To help customized machine studying, the Android ML group is actively growing Android’s customized ML stack. Final 12 months, TensorFlow Lite and GPU delegates had been added to the Google Play Companies which lets builders use TensorFlow Lite with out bundling it to their app and gives computerized updates. With improved inference efficiency, {hardware} acceleration can in flip additionally considerably enhance the consumer expertise of your ML-enabled Android app. This 12 months, the group can also be asserting Acceleration Service, a brand new API enabling builders to select the optimum {hardware} acceleration configuration at runtime. It’s now in public beta and builders can study extra and get began right here.

To study extra, watch the video: