Machine Learning for ABR in production
summary
YouTube integrates machine learning into ABR incrementally rather than replacing it, deploying a v1 in production serving v1 at ~1 Gbps and targeting terabyte‑scale egress. Workflows use TensorFlow as configuration language and train ML on isolated subfeatures; Kaplan–Meier survival curves are proposed for better session metrics. A desktop experiment capping defaults to 480p default caused no engagement loss for most users, and a mobile personalization model saves about 25% egress while producing single‑digit percentage engagement gains.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:35Machine Learning
- 7:45Audience Participation
- 9:44Sticky Quality
- 12:52Mobile Quality
- 15:40Does it work
- 18:29Everyone needs quality
- 19:16Quality is subjective
- 20:04What we dont do
- 21:46Quality selector
related talks
19mReinforcement learning for ABR
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
21mLearning in situ: a randomized experiment in video streaming
22mProviding better video experiences for the next billion users.
20mReconstructing proprietary video streaming algorithms
29mEngineering Machine Learning Technologies
25mML model from an idea to production with the help...
16mImproving Streaming Experience with Bayesian Optimization, from AB to...