Adoption of machine learning has increased dramatically over the past 10 years. What was once a vibrant research field has moved into the marketplace and now is revolutionizing industries such as automotive, manufacturing, agriculture, games, and entertainment. 

As ML moves into production systems, several challenges arise, including how to properly evaluate models for quality and fairness, how to serve model predictions efficiently to satisfy latency and throughput requirements, and how to ensure that model integrity in production environments is not compromised over time. 

That’s where MLOps comes into play. 

What Is MLOps?

MLOps is a set of practices that focus on training, deploying, and monitoring ML models in production environments. 

A key concept that underpins MLOps is that ML models need to continuously evolve to remain highly performant in dynamic production environments. This means that many aspects related to managing models need to be automated to create a tight iteration loop where models are retrained when new data becomes available or their performance begins to degrade.

However, because MLOps is a new field, not much material exists to help familiarize people with it. I’ve selected four online courses that provide a great introduction to MLOps and, more broadly, managing ML in production. One caveat: This is a highly practical field, so it is important to view these courses as a way to jump-start or supplement your knowledge of production ML, not as a replacement for hands-on experience.

Each course provides a solid background in how to design an ML production system end to end, touching on the core concepts of the ML lifecycle. This includes:

• Project scoping (i.e., when to use ML)
• Managing data (i.e., storage, processing, and validation)
• Training models (i.e., deep learning libraries, experimentation, hyperparameter tuning)
• Model deployment (i.e., cloud hosting and serving)
• Monitoring (i.e., tracking data drift) 

The courses differ in how they are structured (lecture notes versus videos), whether they offer a hands-on practical component, the depth in which they cover each of these topics, and whether they cover additional topics such as bias and fairness in ML.

Prerequisites for all four courses are similar, including a working knowledge of AI and deep learning, intermediate Python skills, experience with any deep learning framework (such as TensorFlow, Keras, or PyTorch), and proficiency in basic calculus, linear algebra, and statistics.

All of the courses, except MadeWithML, also require experience with code versioning, Linux or Unix development environments, and software engineering. While some of the courses cover core deep learning topics such as backpropagation, convolutional and recurrent neural networks, and transformers, they are intended to serve as reviews and not a thorough introduction.

Machine Learning Engineering for Production (MLOps) Specialization 

The four courses in this series, offered by DeepLearning.AI through Coursera, are taught by four instructors, including Coursera co-founder and DeepLearning.AI founder Andrew Ng. Coursera also “highly recommends” completing its Deep Learning Specialization before starting these courses. (That said, I don’t think it’s critical to complete this work before pursuing the MLOps specialization.)

Cost: $49/month after a seven-day trial

Time commitment: Approximately four months at a suggested pace of six hours per week; cumulative duration of video lectures: 20 hours

Hands-on practice: There are 20 labs associated with the four courses and several ungraded, optional labs, some of which are available to the public on GitHub.

Last updated: January 2022 (based on the commit log of public GitHub repo)

Upshot: Of the four courses covered here, this one offers the most structured, hands-on assignments, with multiple quizzes, as well as practical application labs. Notably, it also goes deeper into the infrastructure aspects of MLOps, such as Kubernetes.

Full Stack Deep Learning

Offered through the University of California, Berkeley, this course is taught by Sergey Karayev, head of STEM AI at Turnitin and co-founder of Gradescope; Josh Tobin, founder and CEO of a stealth mode startup and a former research scientist at OpenAI; and Pieter Abbeel, a UC Berkeley professor and co-founder of Covariant.ai, Berkeley Open Arms, and Gradescope.  

Cost: Free

Time commitment: 16 hours of video lectures; the nine labs require about 10 hours to complete

Hands-on practice: Nine labs providing a cross-section sample of practical experience

Last updated: Spring 2021

Upshot: The course covers several topics beyond MLOps, ranging from foundational deep learning material, to tips and tricks on how to effectively debug model training, overviews of exciting research frontiers, ML at startups, and the relevance of a Ph.D. Furthermore, the course offers a large number of guided hands-on labs. However, because only about half of the course focuses on MLOps, it provides fewer details than either the Coursera specialization series or the Stanford course described below. 

CS 329S: Machine Learning Systems Design

Offered through Stanford University both online and in-person, this course is taught by Chip Huyen, co-founder of a streaming-first platform for ML, author of the upcoming book Designing Machine Learning Systems: An Iterative Process for Production-Ready Applications, and developer of ML tools for Nvidia, Snorkel AI, Netflix, and Primer. 

Cost: Free

Time commitment: There are no prerecorded lectures, but 16 twice-weekly, 90-minute lectures (24 hours), plus time required to review slides and/or read lecture notes as they’re posted

Hands-on practice: None; end-of-semester project for course attendees (Note: At the end of the semester, project presentations can be viewed by anyone) 

Last updated: Spring 2022

Upshot: Taught on Stanford’s campus but accessible remotely, the course features extensive slides and lecture notes on its syllabus page. Notes include draft chapters of an upcoming book by the course instructor. The course also includes frequent contributions and tutorials from outside experts.

MadeWithML.com

Offered independently by Goku Mohandas, the founder of MadeWithML.com, this is less a “course” than an extensive site of ML educational resources. Mohandas—who conducted a tutorial for the Stanford course—has worked on ML at Apple, startup Ciitizen, and his own startup, HotSpot. 

Cost: Free

Time commitment: Roughly eight hours, based on a cumulative 70,000-word count and reviewing all sections at about 150 words per minute; there are no video lectures or slides

Hands-on practice: Guided projects via notebooks shared on Mohandas’ MLOps GitHub repo, but there are no structured assignments

Last updated: January 2022 (based on the commit log of MLOps GitHub repo, but in theory this course is continuously updated)

Upshot: MadeWithML offers the most gentle introduction to MLOps, since it requires the fewest prerequisites. It includes introductory topics such as Unix, Git, CLI, and RESTful API. Additionally, it has a section dedicated to introducing ML and deep learning. “We never jump straight to code,” Mohandas explains on the homepage. “Instead, we develop an intuition for the concepts first.” Continuously updated material means never having to contend with obsolete information, and the site offers a wealth of well-written notes complete with embedded code snippets. While not as deep a treatise on MLOps as the other courses listed here, it opts for breadth over depth.

Jump-Start Your MLOps Knowledge

ML is revolutionizing industries across the board, and employers want engineers and stakeholders to understand how models should be developed and maintained in production environments. These four MLOps courses can jump-start understanding for anyone from an enthusiastic programming amateur to a seasoned software professional. Think of this overview as a starting point for choosing a course best suited to your individual needs.