Fire Up Your Smartphone Neurons with On-Device Deep Learning
summary
An Android demo pipeline for on‑device deep learning: train with TensorFlow using transfer learning (MobileNet), export a .pb and label file, serve models via REST and download them to phones. Uses MobileNet and post‑training quantization to shrink models from ~80MB to ~20MB, trading slight accuracy loss for much smaller size, faster runtime and lower CPU/memory; benchmarks show single‑frame inference around 2–3 seconds on a Nexus‑class device and reduced power/CPU with smaller MobileNet variants. Also presents a prototype for fusing shared layers across models to save resources and a serving flow for remote model updates.
AI-written summary · about
chapters (18)
- 0:00Audience Introduction
- 0:50Topic Introduction
- 3:32Context
- 5:22First Example
- 6:22Second Example
- 7:44Third Example
- 10:57Fifth Example
- 14:30Seventh Example
- 15:59Eighth Example
- 19:50Ninth Example
- 22:03Tenth Example
- 27:04Eleventh Example
- 35:25Aggregations
- 36:55Model Doubling
- 38:20Model Performance
- 40:50Retrieving Weight Objects
- 42:15Feedback
- 45:20Q&A
related talks
15mSimplified Machine Learning on mobile, browsers, and edge devices with on-device ML
41mWarm Up Your Smartphone's Neurons with Deep Learning
39mAI/ML on mobile/android with TensorFlow
19mOn-device machine learning: TensorFlow on Android
39mCreating smaller, faster, production-worthy mobile machine learning models for Android
39mMachine Learning on Android Demystified
38mNEURAL NETWORKS ON YOUR MOBILE PHONE. DIVING INTO THE NN API
47mMachine Learning on Android Demystified