Learning About Machine Learning: a Primer on MLOps
- Zephin Livingston
MLOps are a fast-growing piece of the AI puzzle, but if you’re still not sure what it is, we’ve got you covered.
MLOps has become a hot-button talking point in the tech industry as of late, for how it can potentially shorten development life cycles and allow businesses to see a faster return on investment from their machine learning projects. The nascent industry is growing at quite a clip: the machine learning company, Neuro Inc., predicts MLOps will generate $4 billion in revenue by 2025.
That’s great, but what really is MLOps? Who is behind its surge? How does it affect the enterprise? Here, we answer those pressing questions in the following primer.
What Are ModelOps & MLOps?
First, some terminology. MLOps is short for “machine learning operations.” It isn’t some new technology or hardware that makes machine learning better; it’s a set of practices. While there are technologies associated with MLOps, such as Microsoft’s Azure Machine Learning, Paperspace’s Gradient, and the open-source MLflow, these are simply platforms designed to better facilitate machine learning. MLOps itself is built on the idea of managing the life cycles of a business’s various machine learning models to keep it all running optimally.
To make things more complicated, if you’re researching a job in MLOps, you’ve probably also come across a similar term: ModelOps. Harish Doddi, CEO of AI platform builder Datatron, differentiates the two terms like this: “MLOps is a subset of ModelOps. Whereas ModelOps focuses on all AI and decision models – including monitoring and governance of the models – MLOps focuses specifically on machine learning models.” Essentially, MLOps is a machine-learning focused set of practices whereas ModelOps is a more generalized set of practices for all AI applications.
Why Use Machine Learning & MLOps?
The economic benefits of MLOps can be hard to define if you’re not someone immersed in the business side of AI platforms. Luis Ceze, co-founder of Seattle-based OctoML, is immersed in this world, and has some insights on how MLOps can help a business grow and expand: “The benefit of using MLOps is to create a systematic flow across these phases and automate the process as much as possible, bringing shorter time-to-value and lower costs.” In other words, it makes the process of using machine learning models in your business more efficient and lets you see the benefits of using those models faster.
Who Are the Major Players in MLOps?
Being such a new field, MLOps counts both tech stalwarts and newcomers among its most prominent firms. Many of these platforms are open source, which makes sense. According to a 2020 Red Hat Report, 90% of IT leaders use open-source software in one way or another. Major open-source platforms include MLFlow, Metaflow, Iguazio, and Kubeflow. Titans of the industry like Microsoft, Amazon, IBM, Alibaba, and Google also have their own MLOps platforms and solutions available like Azure, SageMaker, and Vertex AI and, as expected, command a sizable share of the market.
However, if you’re talking about MLOps, you’re really talking about smaller companies and startups. The PNW has some great companies, including Algorithmia (recently acquired by DataRobot), OctoML, and WhyLabs. Out in the rest of the world, names like Datatron, CometML, Weights & Biases, Snorkel.ai, Seldon, DataRobot, Cloudera, Daitaku, Modzy, and Domino are prominent players. Their work in providing cloud-based platforms, developer tools, software solutions, and analytics make them some of the most exciting names in MLOps
On to the Future
As with any field in tech, MLOps is constantly striving for innovation, for the next solution or tool that will optimize their clients’ models even further. For OctoML’s Ceze, that solution is automation, in particular, he says, “Using ‘ML for MLops’ — as in automatic data labeling; anomaly detection; and automatic model optimization, tuning, and deployment.” Essentially, automation would allow for MLOps’ practices to be streamlined and require less human oversight to be effective.
Datatron’s Doddi believes the “next big thing” will be AI model governance for live production. “In an AI governance solution, the catalog needs to be able to keep track and document the framework where the models are developed,” the CEO says. “The catalog will also need to be able to ensure model lineage where it associates the models with the functionality features within the models. And ultimately, it enables computation of the proper governance metrics of the different features.” In plainspeak, AI governance solutions could allow companies to keep their machine learning models more cohesive and consistent — and the more consistent a machine learning model is, the more value it can bring to a business.
The youth of the MLOps scene are also seeing their fair share of success. In November 2021, Comet received $50 million in funding, just mere months after raising $13 million in an earlier funding round. South Korea’s Nota also managed to snag $14.7 million in funding in November. They’re all always hiring software engineers, software developers, data scientists, and system administrators to help this young sector grow.