Running models locally.
WHAT ON-DEVICE MACHINE LEARNING PROVIDES
Processing without sending data anywhere Working without connectivity No per-request cost Lower latency
WHY THAT MATTERS LOCALLY
No data consumed, and it works when the connection does not.
WHAT THE READY-MADE CAPABILITIES INCLUDE
Text recognition from images Barcode scanning Face and object detection Language identification and translation Digital ink recognition
WHAT THOSE SUIT
Common tasks, without training anything.
WHAT CUSTOM MODELS REQUIRE
A trained model converted to a mobile format.
WHAT CONVERSION INVOLVES
Reducing size and precision so it runs on a device.
WHAT THAT COSTS
Some accuracy.
WHAT TO MEASURE
Accuracy after conversion, and inference time on a modest device.
WHY A MODEST DEVICE
Inference time varies enormously with hardware.
WHAT TO CONSIDER
Application size, since bundled models add substantially.
WHAT TO DO ABOUT THAT
Download models after installation, where practical.
WHAT TO HANDLE
Model download failing Inference taking longer than expected
