Скачать или смотреть Yordan Zaykov: "Probabilistic programming in production with Infer.NET"

"Probabilistic programming is becoming increasingly popular and is now widely used for data analysis. This is made possible by frameworks which abstract inference away and allow data scientists to focus on statistical modelling. However, it can be challenging to go beyond data analysis and incorporate probabilistic programming into real-time systems in business or consumer products. In this talk we will review some of the probabilistic programs that we have shipped in production and analyze what framework capabilities enabled their productization.

All probabilistic programs were written using the Infer.NET framework and integrated with various Microsoft products, in collaboration with each corresponding product group. These models are Clutter – a personalized email classifier, TrueSkill – a skill rating system for online video game matchmaking, Matchbox – a general-purpose recommender, and Alexandria – a system for unsupervised knowledge base construction.

We observe that the two most important Infer.NET features for productizing the above are efficient offline inference and support for online Bayesian inference. This is because many enterprise problems operate on massive data scales and require user personalization at runtime. It is also important that the framework enabled rapid model development and could handle a sufficiently broad range of models."