On-Device Machine Learning with Google AI Edge
summary
A practical walkthrough showing how to run on‑device inference with MediaPipe Task with TensorFlow Lite, covering image classification, object detection and pose/hand/face landmarks (pose 33 points hand 21 face 468) and about 20 built‑in tasks. Shows training with MediaPipe ModelMaker (example: custom object detector for wooden toys, Rock‑Paper‑Scissors gestures), deployment on Android and Raspberry Pi 4B, and conversion/quantization that cuts model size roughly 4x. Demonstrates image diffusion (~10s on a Galaxy S24) and Gemma 2B LLM on-device with tokenizer, sampler and async APIs.
AI-written summary · about
chapters (19)
- 0:00Google AI Edge intro
- 1:10What is Google AI Edge
- 2:53On-device ML benefits
- 4:00MediaPipe Tasks evolution
- 6:08Image classification with TF Lite
- 8:56Image classification with MediaPipe
- 12:10MediaPipe Tasks setup
- 16:19Object & landmark detection
- 19:48Segmentation & face landmarks
- 20:32Custom models with Model Maker
- 23:59Retraining for gestures
- 25:01PyTorch to TFLite conversion
- 27:09Generative AI on device
- 31:22Guided image generation
- 33:25Customizing with LoRA
- 34:47LLMs with MediaPipe LLM
- 39:01How LLM Inference works
- 40:25Implementing LLM Inference
- 41:08Resources & feedback
related talks
40mOn-Device Machine Learning for Android with MediaPipe
32mAdvanced on-device ML made easy
48mImage Recognition and Object Detection with MediaPipe on A…
15mSimplified Machine Learning on mobile, browsers, and edge devices with on-device ML
08mDemo: Gemma on-device with MediaPipe and TensorFlow Lite
23m2025: On‑Device AI Goes Mainstream on Android
31mBringing AI on-Device: From Cloud to Edge
50mML and AI at the edge with Google Coral, EdgeTPU and Cloud Platform