In addition to being a thriving academic discipline, machine learning is increasingly adopted as a solution to real world business problems. But underneath this seeming success lies a chasm of failures. In fact, according to recent surveys among industry practitioners, the majority of ML projects still fail. It turns out that deployment workflow of ML is far from trivial, and adds a pile of its own challenges on top of those that already exist in software development practice.

In this project we study existing reports of ML deployment, and investigate issues practitioners face at each step of the ML deployment pipeline, from data collection and model training to quality assurance and monitoring, as well as cross-cutting aspects such as security and ethics. We show that issues and challenges can affect every stage of the ML deployment, and illustrate them with examples from different fields and industries.

This project is a collaboration with Raoul-Gabriel Urma from Cambridge Spark.

Related Publications

Challenges in Machine Learning Deployment: A Survey of Case Studies

Andrei Paleyes, Raoul-Gabriel Urma, Neil D. Lawrence

ACM Comput. Surv., Association for Computing Machinery: