TinyML: Machine Learning’s Future is Tiny and Bright
summary
Argues TinyML runs on-device at under 1 mW and on kilobytes of RAM, not megabytes, enabling always‑on sensor analytics; demonstrates a 2 KB runtime kernel and interpreter-based deployment where interpreter overhead is ~4.1% so models (16–275 KB) remain portable and updatable OTA. Shows tooling: TensorFlow Lite Micro for MCU inference and a standard benchmark suite MLPerf Tiny; and an automated dataset pipeline that produces 6,000+ hours across 50 languages, proving practical multilingual keyword spotting at low power.
AI-written summary · about
chapters (26)
- 0:00Intro
- 2:15Terminology
- 3:07Applications
- 3:59Detection
- 4:47Segmentation
- 5:28Machine Translation
- 8:02Computational Demand
- 9:52Historical Trends
- 10:25Big Compute
- 12:40Tradeoffs
- 13:18Energy Consumption
- 16:03Edge Computing
- 16:52Image Enhancement
- 17:26Sign Email
- 17:46Why TinyML
- 22:46Summary
- 23:52Machine Learning Capabilities
- 24:20What is TinyML
- 29:50Use Cases
- 31:02Autonomous Drones
- 32:26ARVR
- 33:49TinyML on Device
- 34:20Whats Next
- 36:49Embedded Systems
- 44:05Challenges
- 52:14Tensorflow
related talks
27mWhat’s TinyML good for
25mMachine Learning on Microcontrollers; Tiny, Useful and Fun
31mKeynote: The next big thing in machine learning is tiny
25mTinyML as-a-Service
36mTinyML at the Edge: Deploying and Optimizing AI Workloads on Zephyr RTOS
01h 01mML and the IoT: Living on the Edge
32mNext level Machine Learning with TinyML and Python
10mMachine Learning for mobile