Self-learning Alexa: ML model updates with no human in the loop
summary
A real-time continual learning system labels live Alexa audio with semi‑, weak‑ and self‑supervised methods (teacher→student distillation, noisy‑student, representation learning), computes gradient updates in 30-minute rounds, sends aggregated updates from hundreds of ECS hosts to an orchestrator, then discards raw audio so models improve without persisting user audio. Results show about a 2.7% relative WER gain overall, ~20% improvement on new words and phrases, and <1% degradation after adding memory replay; large batches let incremental training converge in days rather than weeks.
AI-written summary · about
chapters (34)
- 0:00Intro
- 1:04Alexa recognizes speech anywhere Device location
- 2:13Alexa's growing footprint
- 3:05Alexa recognizes speech for anyone
- 3:46An example: pajamas
- 4:25What is self-learning ML?
- 6:22Speed: Trending topics
- 7:07Privacy
- 7:21Personalization
- 7:58What is self-learning?
- 8:56Semi-supervised learning
- 11:18Weakly supervised learning
- 13:31Self-supervised learning
- 15:55Incremental learning
- 17:50Traditional ML
- 18:27Real-time self-learning
- 19:19Inspired by federated learning
- 22:00The system at 10,000 feet
- 23:26Learning service
- 23:55Learning sources
- 24:42Selecting training samples
- 26:59Labeling the training samples
- 27:46Model update calculation
- 29:15Learning orchestrator
- 29:57Updating the global model
- 31:36Mixing in offline data
- 32:33Evaluating the updated model
- 33:37Rinse and repeat
- 34:01Alexa ASR results
- 35:02Improvement on recent data
- 37:13Large batch size to reduce training time
- 39:36More takeaways
- 41:29Next steps
- 43:17Summary
related talks
32mHow Alexa keeps getting smarter
39mIt's All in the Data: The ML Behind Alexa's AI Systems
54mMachine Learning for Developers
50mBuilding Continuous Learning Systems
55mMachine Learning and AI in Java on AWS
29mEngineering Machine Learning Technologies
18mUsing Machine Learning to Enrich and Analyze Real-time Communications
31mLearning the learner: Using machine learning to track performance