NEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
summary
Overview of running trained neural nets on-device with concrete tradeoffs: Android 8.1 NN API exposes native C paths that already give ~3–4x speedup on CPU and route work to accelerators like the Kirin 970 NPU when present. Use TensorFlow Lite (TF Lite) with NNAPI for best performance but expect model conversion issues and sometimes needing to change ~50% of a model; models commonly weigh 40–300 MB. ML Kit covers many common features locally; beware ambiguous outputs and sensor‑based privacy leaks.
AI-written summary · about
chapters (15)
- 0:00Introduction to talk
- 1:02Android and AI background
- 3:53Training models on edge devices
- 6:03Why neural networks now
- 8:59Word vectors and translations
- 12:11Edge computing benefits
- 15:21TensorFlow for mobile
- 18:29Neural Network API overview
- 21:41Real-time classification demos
- 25:41Deeper classification models
- 26:23ML Kit for mobile AI
- 30:43Interactive neural network demo
- 33:15KAS and model setup tools
- 34:44Sensor data and privacy
- 37:16Resources and conclusion
related talks
39mAI/ML on mobile/android with TensorFlow
46mFire Up Your Smartphone Neurons with On-Device Deep Learning
39mMachine Learning on Android Demystified
06mPyTorch Mobile and Android Neural Networks API
23mAI is ready for you. Are you ready for AI?
47mMyths and Reality of Android Machine Learning
40mApplied TensorFlow in Android apps
32mAI for Mobile and IoT Devices: TensorFlow Lite