Engineering Machine Learning Technologies
summary
Claims that open source + rapid iteration turns research into deployable products, citing 6 months from research to product for Smart Reply and an engineering path that mixes signal processing, training trickery and systems limits. Uses Google home 4-mic beamforming with ML-estimated filter parameters (an AutoML-like loop), trains on simulated noisy mixes from YouTube, combines acoustic and language models with multi-head attention, and leverages pretrain with BERT and fine-tune to cut data needs while checking CPU/power for on-device deployment.
AI-written summary · about
chapters (10)
- 0:00Start
- 0:12Motivation
- 2:25Open Source Paradigm to the Development of Machine Learning
- 3:28Automatic Replies
- 8:28Ai Principle
- 13:14Speech Recognition
- 15:11Neural Data Beamforming
- 20:42How To Train the Models
- 24:49Natural Language Processing
- 27:31Moral Risks
related talks
47mBuilding smarter apps with Machine Learning, from magic to reality
01h 03mUsing Machine Learning and AI to serve people and communities
07mMachine learning at Google
47mMyths and Reality of Android Machine Learning
44mBuilding smarter solutions with no expertise in machine learning
57mBuilding smarter solutions with Machine Learning, from magic to reality
14mMachine Learning for developers in a hurry
54mBuilding smarter apps with Machine Learning, from magic to reality