MLOps operationalizes the production of machine-learning software so agencies can expedite the continuous production of ML models at scale, significantly reducing the time it takes to deploy intelligent AI applications.
Just when government agencies were getting accustomed to integrating DevOps practices into their workflows, a new “Ops” has emerged: MLOps, a practice Deloitte has named one of the key government technology trends for 2021. So what is MLOps, and why should government agencies care about it?
MLOps brings together data scientists, developers and IT operations professionals to efficiently deliver machine learning models into production. Consider MLOps a subclass of DevOps: It extends and applies DevOps concepts to ML code, improving the quality of artificial intelligence-powered applications and reducing the time it takes to go from idea to deployment.
That’s important for government agencies investing money into AI-related R&D. With the federal government predicted to invest more than $6 billion in AI in fiscal year 2021, agencies will want to ensure their investments bear fruit. That’s where MLOps comes in.
How MLOps works
The MLOps program is similar to traditional DevOps. Each team member has a role to play in the production pipeline. For example, data scientists prepare data, apply ML algorithms and tune the models to make them more performant. Developers use those models as part of their applications, and operations managers ensure that models are approved and monitored in production.
As with traditional DevOps, the goal of the MLOps deployment pipeline is to streamline the delivery of software into production, with the added twist that the software can make predictions based on machine-learned patterns in historical data. MLOps encourages automation and continuous deployment just like DevOps, but it adds unique ML capabilities such as model validation to enable high-quality predictions to be made in the environment.
The work of data scientists is often referred to as “experimentation.” Data scientists test and retest algorithms on data and iterate on the result to improve performance. While experimentation can lead to powerful ML solutions, without a deployment pipeline, those solutions will likely stay in the lab.
MLOps operationalizes the production of machine learning software and enables agencies to benefit from the experimental efforts of its data scientists. It provides a foundation for ML to have tremendous impact and value.
Delivering viable solutions
Meanwhile, MLOps also gives agencies a methodology they can use to maximize AI investments and deliver viable solutions. By adopting MLOps best practices, agencies can expedite the continuous production of ML models at scale, significantly reducing the time it takes to deploy intelligent AI applications to days or even hours.
The potential use cases are eye opening. For instance, agencies can use MLOps-built models to predict equipment failures on military vehicles and proactively develop appropriate maintenance schedules. They can train computer vision models to analyze satellite or astronomical images, or they can develop truly innovative solutions that allow for edge compute processing in minutes.
MLOps also injects repeatability and auditability into the deployment pipeline, which can be beneficial if an ML model underperforms or an AI “black box” needs to be audited post-deployment. Having a proper deployment pipeline in place makes it easier to regenerate models and retrace steps if necessary.
Getting started with MLOps
From an operational point of view, deploying an MLOps program is similar to the creation of a typical DevOps program. The first step is to create a collaborative culture with cross-functional teams working to meet their agency’s business goals. Then, that culture must be supported with the appropriate underlying technology to help team members work together and achieve those goals.
Fortunately, government agencies with DevOps teams have already laid much of the cultural groundwork necessary for MLOps; they simply need to add their data science teams to the mix. And while data scientists are typically used to working alone, they will benefit from hearing conversations around product requirements and getting to know their agencies’ developers and operations managers. These insights can help data scientists align their models with the goals of the product and help developers and operations managers understand how those models were built and meant to be consumed.
Teams should be working toward a common goal that aligns with their agency’s primary AI objectives. Depending on the agency, that goal could be centered on their use case, such as predictive maintenance for more effective medical treatments. The goal should not be optimizing a specific engineering metric, but a strategic organizational objective that MLOps teams design applications to meet.
Finally, agencies should support these cultural and strategic initiatives with technology that helps MLOps teams to easily build and deploy ML-driven applications. The ideal platform should support information sharing across teams and enable agile development processes. Open source platforms enable data scientists, developers and operations managers to use different tools while still facilitating collaboration.
More than just another “Ops”
While MLOps is a variation on what has become a very popular theme, it’s much more than just another “Ops” methodology. It can help take ML out of the experimental phase and into practical use. It not only brings data scientists, developers, and operations managers together. MLOps has the potential to bring government agencies’ AI projects to life.