Myths and Reality of Android Machine Learning
summary
Argues that modern Android runs practical on‑device ML using NNAPI and ML Kit and shows live demos: object detection feeding product search, on‑device Translate (59 languages) and smart replies with on‑device privacy. Reports concrete gains: 9x latency reduction (face detection, ~495 ms → 55 ms), 3.7x energy reduction for OCR. Deploy options are APK, Firebase model serving or dynamic module. Recommends combining camera preview with analysis to avoid lag and using AutoML Vision Edge with about 50 images per class to prototype.
AI-written summary · about
chapters (32)
- 0:00Introduction
- 1:53Android is not powerful enough
- 3:15Different stages of ML
- 5:34How to package ML
- 6:34Demo
- 8:22Results
- 9:28Technical Improvements
- 11:08ML Code
- 11:47Top 1 Results
- 13:46People AI Guidebook
- 16:03Material Design
- 17:24PhDs
- 18:02ML Kit
- 18:13Language Identification
- 20:15Market Price
- 23:16User Engagement
- 23:37Google IO
- 24:16How does it work
- 31:14API level 8
- 31:52How is it done
- 35:42Another way
- 36:15Google Collab
- 40:28Machine Learning Needs Data
- 41:25ML Kidotto
- 41:40Vision Edge
- 42:03Busy
- 42:26Compute
- 42:42Enterprise Environment
- 44:38Sample Size
- 45:06Machine Learning is not a solution
- 45:49Recommendation
- 46:38Questions
related talks
43mMyths and Reality of Android Machine Learning
43mWhat’s new in Android machine learning
39mMachine Learning on Android Demystified
13mWhat's new in Android machine learning
36mMaking intelligent Android apps
30mMachine Learning in Android Apps with ML Kit
39mAI/ML on mobile/android with TensorFlow
47mMachine Learning on Android Demystified