Software Engineering Radio - the podcast for professional software developers SE Radio 660: Pete Warden on TinyML
Mar 18, 2025
Pete Warden, CEO of Useful Sensors and a founding member of TensorFlow at Google, delves into TinyML—machine learning for low-power devices. He discusses its real-world applications, from voice activation to offline translation, and emphasizes the importance of local processing for privacy. Warden shares insights on challenges like model compression and deployment. He also highlights its potential in agriculture and healthcare, advocating for practical approaches for beginners eager to dive into TinyML development.
AI Snips
Chapters
Transcript
Episode notes
TinyML vs. IoT
- The Internet of Things (IoT) hasn't fully succeeded because simply adding connectivity doesn't guarantee innovation.
- TinyML prioritizes making individual objects smarter locally without internet connectivity.
TinyML Inspiration
- At Google, Pete Warden learned about 30-kilobyte machine learning models used for wake word detection.
- This sparked his interest in TinyML and its potential for running ML on cheap, low-power devices.
Data and Training in TinyML
- TinyML models require substantial data for training, sometimes even more than larger models.
- However, training is more accessible due to lower GPU requirements because of the model's small size.