Many machine learning (ML) projects stall between proof-of-concept (POC) and full-scale production. Successfully scaling pilots across the organization requires a combination of business leadership buy-in, the right skills and data, and ensuring that models remain accurate over time.
“A lot of customers get excited with pilots, but the reality and where they actually benefit is when those POCs become real solutions in production,” says Sri Elaprolu, senior leader, Amazon Machine Learning Solutions Lab.
One barrier to scaling pilot projects occurs when a team fails to get executive buy-in from the business at the start. “POCs can be started by small teams, but we’ve seen a POC get stuck because the executive or business teams were not included in the discussion during the early stages,” says Elaprolu. In those cases, he explains, “when the technical team goes to the business team after the POC and says, ‘Here’s a great solution,’ the business team may not accept that.”
A better way to get business buy-in is through a “working backwards” concept, where teams first agree on the business outcome they wish to achieve, and then design the solution and approach to best achieve that outcome.
This allows the team to determine the use case requirements needed to achieve the outcome, and which data sources will be needed, before putting the technical building blocks in place.
A second key is to make sure your team has the right skills to conceive, build and maintain ML models. This includes highly specialized data scientists of course, but also analysts who can translate business problems into ML solutions. Finding these skills is easier said than done; in the 2020 State of the CIO study, data/analytics ranked #2 on IT leaders’ list of most difficult skill sets to find, just behind cybersecurity. To help close the talent gap, Elaprolu suggests upskilling existing teams where possible to transform them into data-oriented teams.
Another way to maximize limited analytics talent is to leverage partners to remove what Elaprolu calls “undifferentiated heavy lifting” from in-house staff, so they can focus on higher-value activities.
The right data – and enough of it
A third and possibly the most critical factor to scaling ML projects is to make sure the team has the right data, and enough of it.
“It’s easy to get started on a POC because by definition, it means we are looking at a smaller scope problem and we are trying to solve that,” says Elaprolu. “Therefore, we can get that going with a limited amount of data. But in the real world, when you take the solution and deploy it to production, you may run into data conditions that are not necessarily the same as what you’ve dealt with in the POC.”
ML development is often mistakenly grouped with software development’s traditional “set it and forget it” approach. “Software development is based on defining a set of rules to follow, and unless the rules change, you don’t need to change the software,” says Elaprolu. “Machine learning is the opposite: You start with the data and you allow machines and algorithms to learn based on the data, which means you can’t just build, deploy and forget, because you risk concept drift.”
That’s why iterating on ML models is an ongoing process, since conditions and business environments often change in the real world. “You need to have that continuous loop of retraining and redeploying,” says Elaprolu.
Following the three best practices described here will go a long way to turning a POC into a production-ready model that can scale across the business.
Learn more about machine learning and AI services for your business.
For more machine learning insights from Sri Elaprolu, check out the new Ahead of the Pack podcast.