CI/CD for Machine Learning - Detroit 2018

As we venture into new fields, we sometimes forget to apply the lessons learned in the past. As machine learning (ML) becomes more accessible to developers without a data science degree, and ML models multiply, we begin to see a glaring deficiency.

Many of the ML frameworks support easily building and tweaking ML models, but do not offer source control. To make matters worse, ML models often require prolonged trial and error, and even small changes in parameters can produce large changes in model accuracy. As developers test new models, it becomes difficult to roll back to previous versions, create a reliable CI/CD pipeline into the production environment and monitor the application after it is deployed.

In this talk, I will discuss the potential ways you could begin with the end in mind and build a reliable process for delivering your ML models into production.