MLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
summary
Prediction: mobile video rises to 75–80% of mobile data within five years. Practical fix for buffering is adaptive bitrate streaming that prefetches and lowers bitrate based on learned user routes (example: daily 9:30am flyover), with reinforcement learning experiments like MIT’s Pensieve near production. Video indexing requires 24–60 frames per clip, CNNs still lag image accuracy (images ~97%), and one published ABR variant reports about 25% improvement in QoE. Akamai and ad/CDN systems deploy these models to cut costs.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:30Global Communication Transforming to Video
- 2:40User Expectations
- 4:00Industry Bubble
- 5:30Can AI help
- 7:25Adaptive bitrate streaming
- 10:25Adaptive bitrate algorithms
- 12:10Optimal control algorithms
- 16:40Image classification
- 20:30Ordinance and CDN
- 22:45Targeting
related talks
23mMachine Learning for ABR in production
17mContent Aware Encoding for low latency live streaming encoders using deep learning
21mLearning in situ: a randomized experiment in video streaming
17mPredict, Prune, Play: Efficient Video Playback Optimization Under Device...
21mWhat to do after per-title encoding
17mYour Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
23mML-Powered Search & Recommendation 101
12mMeasuring live video quality with minimal complexity, now available for everyone!