PAPIs Europe 2018

Our industry is rapidly changing and we frequently need to upgrade our product to meet customer needs. This is non-trivial as our system runs 24x7 at huge scale, and uses unattended, automated machine learning to spend serious amounts of money every second.



In this presentation we describe a recent algorithm improvement to exploit a previously unanticipated situation. Although we will describe the algorithm, the presentation is focused on describing the pattern we have developed and perfected to safely upgrade our system and evaluate new algorithms while it is running.

Machine Learning is the cornerstone of ad-tech industry organizations. Building scalable pipelines for real-time bidding is crucial for buyers to compete for the most performing inventory on the Internet. We describe the Machine Learning pipeline that we built at AppNexus that operates globally and at Internet scales. The presented solutions describe a qualitative evaluation of our design choices, related to scalable processing infrastructure, high frequency model updates, and the pitfalls that must be avoided when interacting with low-latency model serving infrastructures for RTB.

Open source machine learning tools to manipulate and store data and apply complex algorithms to this data have matured greatly in recent years. However, the final step in any successful machine learning project is to put the models into production with exposed APIs as well as monitor, scale and continuously update them. This final stage has generally been built in-house by large companies and can be a considerable risk to the success of projects within more resource constrained companies. This talk will review some of the challenges in deploying machine learning models and introduce some open source projects that attempt to solve these challenges.

Is it possible to predict the Oscars? BigML’s Deepnets predicted 6 out of 6 Oscar categories right: best picture, best director, best actress, best actor, best supporting actress, and best supporting actor. But how is it possible to predict a seemingly random event? In this talk, M.Sc. Poul Petersen, Chief Infrastructure Officer at BigML will show how to approach a problem like predicting the Oscars, how to chose the data that is relevant, how to prepare the data, and how BigML's Deepnets work behind the scenes to give the best possible model for your machine learning problem.