DevOps / Machine Learning / Contributed

Industrialize Machine Learning to Minimize Technical Debt

3 May 2021 3:00am, by

Diego Oppenheimer
Diego Oppenheimer is co-founder and CEO of Algorithmia. Previously, he designed, managed, and shipped some of Microsoft’s most used data analysis products including Excel, Power Pivot, SQL Server, and Power BI. Diego holds a Bachelors degree in Information Systems and a Masters degree in Business Intelligence and Data Analytics from Carnegie Mellon University.

We’re in the midst of a major upswing in the adoption of artificial intelligence (AI) and machine learning (ML). More and more organizations have realized just how critical these technologies are to their ability to remain competitive and they’re investing accordingly, with the vast majority of AI and ML budgets and staffs growing.

This was one of the most prominent trends in our 2021 enterprise trends in machine learning report, in which 83% of survey respondents said their AI/ML budgets have increased year-over-year despite economic uncertainty related to the pandemic. The average number of data scientists working in the companies included in our report grew 76%, reflecting a corresponding increase in hiring. These organizations are moving very quickly to capitalize on the top-line and bottom-line opportunities machine learning creates for their businesses, aided by the proliferation of simplified tooling that lowers the barrier to entry. You should be doing this, too, if your own initiatives aren’t already underway. There are virtually no industries where AI/ML can’t generate business value.

While this mainstream momentum is great to see, our report also found a concerning trend that companies must address now with the same level of urgency: Teams get started quickly and move fast, but they do so without a clear sense of direction as to where they’re headed. They begin to deploy models to production without a complete understanding of how these newer technologies integrate with their people, processes, and technology stacks. And they build and operate everything in a highly manual way.

This reflects the creativity and persistence of the data scientist: They’re willing to do whatever it takes to get their models in production, and they’re not afraid to get their hands dirty with operational tasks in order to achieve early wins.

There are virtually no industries where AI/ML can’t generate business value.

But this causes technical debt to rapidly accumulate. Data scientists spend more of their time just getting models deployed, and less time building and iterating innovative models — which is where their true talent lies. The problem worsens as the organization scales its initiatives. The team that seemed to be moving at top speed hits a wall and has to backhaul its operations to fix underlying problems, costing critical time and resources. We’re seeing clear warning signs flashing that this is already a significant issue.

Our independently conducted survey included input from more than 400 leaders and practitioners across a wide variety of technology and business roles, each of whom has a direct stake in their company’s machine learning strategies. We found that even as they significantly boost their ML-related spending and hiring, they end up directing more of these growing resources into manually scaling their initiatives instead of investing in their operational efficiency and generating even greater growth.

The time needed to deploy a trained model to production increased year-over-year even as investments soared, for example; 64% of organizations take a month or longer to deploy a model. Surprisingly, organizations with more ML models running in production spend more of their data scientists’ time on model deployment, not less. At 38% of the organizations included in our report, data scientists spend more than half their time on operationalizing their models — and the more models an organization has in production, the worse these numbers tend to be.

These are eye-opening numbers, and they are not sustainable operating models. Ad-hoc manual processes, disparate teams and tools, and other issues are causing technical debt to balloon to dangerous levels. This in turn limits the amount of business value organizations can derive from their increasing investments in machine learning.

Some technical debt is inevitable. Tradeoffs have to be made to achieve certain business goals faster than would be otherwise possible. Used properly, technical debt is a strategic tool. But just like with financial debt, major problems occur when you let it get out of control and have no plan to pay it back. Too much tech debt will absolutely undermine your AI/ML initiatives.

You need to address this issue now by investing in your operational efficiency and scale. We think of this crucial step as industrializing your ML, and the best way to do this is with machine learning operations (MLOps). MLOps is a discipline that fuses the practice of AI/ML with DevOps principles to give data science, product, and IT operations teams everything they need to rapidly and efficiently deploy, manage, govern and secure their models at scale. This allows businesses to scale their initiatives in a rapid, repeatable manner that reduces the manual effort otherwise required of your teams to simply deploy a model to production. This leads to faster results that generate tangible business value from your machine learning initiatives.

Implementing MLOps also helps you avoid a lot of future technical debt, and it’s never too early to start. It’s useful if you have just one model, and it becomes increasingly valuable as the number of models you’re running in production grows.

Operationalizing models should get easier and easier as you scale, not harder. MLOps makes this possible, both by minimizing technical debt as well as ensuring that your people and processes can operate at maximum speed and efficiency. Speed is of the essence in today’s business world. Industrializing your machine learning with MLOps allows companies to rapidly adapt to changing signals in their data and maximize results.

Companies that industrialize ML are able to act much faster in response to the data-driven insights that their models produce. They gain a significant competitive advantage not because they have the best models but because they deploy their models faster and more frequently, and they’re able to rapidly iterate based on the signals those models produce. Their newfound agility reduces the time they spend on manual operational toil. As a result, they are better positioned to capitalize on new business opportunities and rapidly respond to new risks.

This is why AI/ML is gaining a large foothold in just about every industry. There’s virtually no business where it’s not applicable, from healthcare to financial services to retail and more. Companies are realizing that machine learning applies to so many facets of their respective businesses, such as discovering new revenue streams, boosting customer experience, optimizing their supply chains, and reducing operating costs. Properly operationalized ML models enable them to react to new or changing market conditions as they occur, versus playing catch-up months down the line.

This top-line and bottom-line business value require the speed and agility that companies gain when they industrialize their machine learning. Responding to changing digital signals in your own business depends upon similar speed and agility. MLOps is what makes this possible.

You don’t need to go it alone, either. Just as simplified tooling made it easier to get a first model into production, third-party MLOps tools can help you move with similar urgency toward industrialization. You don’t need to waste precious time by waiting to build it yourself later. In fact, our 2021 report found that organizations using a third-party MLOps solution see improved outcomes compared to those that take a completely DIY approach to their model deployment and management infrastructure. They deploy models faster, they spend less of their budget on infrastructure, and their data scientists spend less time on operational toil.

This last outcome in particular speaks to the people part of MLOps: The benefits can extend throughout your organization. MLOps is sort of like training the rest of the business on best practices, automated processes, and standardized tooling — which is key because AI/ML success depends heavily on interdisciplinary collaboration. And your data scientists get more of their time back to do what they do best: Build innovative, creative models that produce tangible business value.

MLOps is not something you need in the future as you scale — you need it today. Build for speed and agility now, while minimizing the technical debt that will otherwise slow your organization down to a crawl later. It’s great to see so many businesses and industries prioritizing AI/ML. Now bring this same urgency to industrializing your initiatives with MLOps. Doing so will maximize the short and long-term impacts of your spending and hiring, and ensure your competitive edge.

Feature image by Alice Pasqual on Unsplash.

A newsletter digest of the week’s most important stories & analyses.