MLOps can streamline machine learning development, thus increasing operational effectiveness. With this in mind, Joey Jablonski, VP, analytics, Pythian, looked at the journey to defining and implementing an MLOps solution for the organization during his Data Summit 2023 presentation, “Building MLOps Organizations for Scale.”
The annual Data Summit conference returned to Boston, May 10-11, 2023, with pre-conference workshops on May 9.
According to Jablonski, high model performance, elimination of bias, and predictability are all key elements of an MLOps strategy.
“We want to think about building technology platforms for analytical knowledge management,” Jablonski said. “We need to enable teams and people to do what they’re best at.”
MLOps needs an environment strategy that includes bias detection, experimentation tracking, model versioning, and model performance.
“We want to make sure we understand where we see improvement, decline, and enable people to use that information later,” he said.
Google and FaceBook are examples of companies that follow this model. However, those organizations have massive engineering teams and could build tools as the organization scaled. Not every company has these tools at their disposal.
Organizations need to look to vendors to implement MLOps, he explained. These vendors include Microsoft Azure, Wallaroo, Domino Data Labs, and more.
“Think about your own environment as you’re looking at these tools and these vendors,” Jablonski said.
According to Jablonski, data science is iterative, “our first (and probably hundredth) attempt will not meet our business needs.” A MLOps environment must track experiments, the parameters associated with them, and the user base impacted. This is about reproducibility, so organizations can properly align analytical models with business processes.
“Think about incrementally improving your technology, your people, and your process at the same time,” Jablonski said.
According to Jablonski, the four steps for success include:
- Define your operating model
- Identify and build skills
- Start small
- Decide on technology
Many Data Summit 2023 presentations are available for review at https://www.dbta.com/DataSummit/2023/Presentations.aspx.