You’ve probably heard it more than once: Machine learning (ML) can take your digital transformation to another level. It’s a pie-in-the-sky statement that sounds great, right? And while you’d be forgiven for thinking that it might sound too good to be true, operational ML is, in fact, achievable and sustainable. You can get the very kind of ML you need to increase revenue and lower costs. To help teams work smarter and do things faster. To make what’s currently impossible into a daily reality.
Unfortunately, most organizations run into trouble when it comes to bridging the gap that exists between experimentation and full-scale ML production. At Cloudera, we spend countless hours with the world’s largest enterprises understanding where the barriers to successful ML adoption are. We then translate those real-world requirements to practical workflows, products, and best practices that help you quickly get to production and scale AI use cases across your business.
We recently published a Cloudera Special Edition of Production Machine Learning For Dummies eBook. In it, we detail what’s needed to succeed with production ML and how to successfully apply a production ML approach at scale within your enterprise. Chapter six of the eBook focuses on the 10 steps for making ML operational. These steps are absolutely critical to helping you break down barriers across the ML lifecycle, so you can take ML capabilities from research to production in a scalable and repeatable manner. Once you can check these 10 steps off your to-do list, you’re that much closer to achieving operational ML.
Step 1: Take a holistic approach to ML
Adopting the right mindset makes all the difference, and that means having a willingness to take a holistic approach to ML. Before ML can become a catalyst for change, it must first be treated as an integral part of your data strategy. By integrating ML and running it alongside your existing IT environments, processes, applications, and workflows, you’ll drive greater business results.
Step 2: Be willing to experiment and, yes, fail
With ML comes the promise of automating business processes and solving business problems. Still, at its core, ML is about science. Proper science takes experimentation and observation, as well as a willingness to accept the failures alongside the successes. Fortunately, even failures can be viewed as victories; once you find that a specific business problem can’t be solved with ML, that knowledge frees you to focus your efforts elsewhere.
Step 3: Build a multi-disciplined team and don’t box them in
Collaboration and freedom from organizational restriction are key. Your data scientists will want a platform and tools that give them practical access to data, compute resources, and libraries. Furthermore, operational ML works best when it’s developed and maintained by a team comprising a diverse range of skill sets—from data engineers and data scientists to even business stakeholders.
Step 4: Iterate quickly. Optimize later.
Don’t worry about building an ML model that’s flawless from the start. Let your teams experiment rapidly, fail early and often, continuously learn, and try new things.
Step 5: Choose the right technology to optimize your entire ML lifecycle
Your data engineering and data science teams need the ability to work across and control the entire ML lifecycle, which Cloudera divides out into two phases:
- Phase 1 of production ML: This phase covers holistic ML development and the building of ML models.
- Phase 2 of production ML: This phase focuses on getting to production, scaling, and ongoing operations.
The right platform and tools will empower your teams to work seamlessly across both of these phases.
Step 6: Evolve your organization to embrace ML
If you’ve already dabbled in ML, you might’ve noticed there’s a wall that exists between experimentation and large-scale production. This wall is there because your organization may lack the knowledge and skills needed to weave ML development, production, and maintenance into your existing processes, workflows, architecture, and culture. That’s why embracing ML requires flexibility in the structure of your organization.
Step 7: Maintain the integrity of your models
Let’s jump ahead to a future where you’ve successfully deployed a few ML models at scale. That’s fantastic! However, your work with those models is far from over. Why? Because the underlying data driving those models shifts over time. Once you have an effective model in place, keeping it fine-tuned takes continuous effort.
Step 8: Close the skills gap
Try to build a team with experience, talents, and capabilities across a wide range of skill sets—everything from data engineering and data science to DevOps and product development. The more diverse this team is, the more its members can learn from one another. Our eBook goes into greater detail about finding and hiring ML talent.
Step 9: Treat models in production like living software
In a sense, models are very much alive. As mentioned in Step 7, models must be fed, sustained, and controlled. Additionally, your models need to be protected. And that makes having visibility into model lineage and monitoring who can access and make changes to your models crucial.
Step 10: Understand and abide by your ethical obligations
Trust us on this one. Don’t sleep on Step 10 because there is no shortage of ethical considerations when it comes to ML. For starters, make sure you have consent from customers and other stakeholders before applying the necessary data against a ML model. Establish and adhere to a rigorous set of ethical ML standards early on. Doing so will pay off down the road.
Making ML work for you
As the saying goes, the journey of a thousand miles starts with a single step—and these 10 steps to making ML operational will get you well on your way to successfully scaling AI use cases. Download the Production Machine Learning For Dummies eBook now to learn how to get there with a proven production ML approach.