Linear Digressions Federated Learning
May 8, 2017
Explore the fascinating world of Federated Learning, where algorithms learn from distributed data while ensuring user privacy. Discover how mobile devices, like smartphones, transform into powerful platforms for machine learning, capturing user interactions without compromising security. Learn about the challenges of decentralized, imbalanced data, and how phones send compressed updates instead of raw data to maintain efficiency. Delve into the innovative workflow that protects user information while enhancing features like autocomplete and photo predictions!
AI Snips
Chapters
Transcript
Episode notes
On-Device Learning Solves Multiple Constraints
- Federated learning lets models learn from user interactions on-device instead of sending raw data to a server.
- This approach addresses decentralization, bandwidth, privacy, and device constraints simultaneously.
Teaching App Revealed Data Collection Tradeoffs
- Katie describes building an Android XML teaching app and collecting keystroke data to learn from users.
- She stopped because continuous data collection would disrespect users' bandwidth and privacy.
Mobile Data Challenges Shape Algorithm Design
- Device datasets are decentralized, imbalanced, throughput-limited, and non-IID, requiring robust algorithms.
- Models are often much smaller than the aggregate training data across devices, favoring update-based strategies.