Xavier Amatriain, Director of Algorithms Engineering, Netflix
summary
Netflix runs ~50 million members, logs ~100 billion events per day and reports 7 billion hours in Q2, so engineering choices matter as much as algorithms. Ten concrete lessons include using stratified sampling instead of throwing all data at matrix factorization, avoiding time traveling features, distributing at three levels (region, hyperparameter sweeps, data-parallel GPUs), training neural nets on GPUs in Amazon Cloud, using multiarmed bandits to correct presentation bias, separating offline item-factor computation from online user solves, and preferring higher regularization within the metric noise.
AI-written summary · about
chapters (8)
- 0:00Start
- 9:28choose your samples
- 10:32try a more complex model
- 12:04give you a couple of examples in the netflix case
- 17:18showing the results of your algorithms
- 22:54optimize a different model for each region
- 23:26pick the best optimal combination of those hyper parameters
- 24:42distributed neural network training on gpus
related talks
22mJustin Basilico, Research/ Engineering Manager, Netflix
41mRecommendations for Building Machine Learning Software
04mInterview
38mNetflix's Ray Platform: From Deep Learning to GenAI
53mKeynote Interview
29mFrom encoding to machine learning a story of managing media data at Netflix
03mInterview
32mRebuilding the AI Platform: Designing for Today’s Chaos and Tomorrow’s Unknowns