Best Practices for Deploying Machine Learning in the Enterprise
Business + ManagementConferencesMachine LearningWest 2018posted by Elizabeth Wallace, ODSC July 18, 2019 Elizabeth Wallace, ODSC
This post discusses the best practices for deploying machine learning in the enterprise.
If you’re an organization worried about being left behind with deploying machine learning, it’s not just you. According to Gartner’s Hype Cycle Chart, machine (and deep) learning are the biggest hyped trends of the year. More businesses, organizations, and startups are talking about deep learning, and what it means for business, than just about any other technological advance.
[Related Article: How to Integrate AI: Digital Adoption Plus Human Resources]
It’s tough, however, because deep learning is still in its infancy. This can make it difficult for businesses that want to implement deep learning models but have no idea where to start. The documentation isn’t there, and use cases are still relatively new.
Robbie Allen, CEO of Infinia ML, has some advice for organizations looking to deploy deep and machine learning in the enterprise. Let’s see the kinds of steps you might need to be successful with such new models.
The Steps for Deploying Machine Learning in the Enterprise
If you’re going to deploy machine learning in your enterprise, you’ll need to keep these three steps in mind:
- data preparation
It all starts with the data. There is no machine learning without quality data. If your data is better than your competitor’s, you have a competitive edge. So think through what’s available to you before even beginning your machine learning initiatives.
- Is the data accessible? Is it even available? For example, sometimes the data you’ve collected over time could be out of touch your terms of service.
- Is it sizable? Do you have enough data? The amount depends on the scope of the problem and the variability of your predictions.
- Is the data usable? Even if you’ve got lots of data, you must be able to use it. You’re spinning your wheels if your data isn’t clean.
- Is the data understandable? Having a proper schema or data dictionary is good to have for making rapid use of data.
- Is the data maintainable? If the hassle for getting the data together is difficult or you don’t even know how the data was put together, it’s a huge problem when it comes to production.
Evaluating Business Opportunities
Getting started with machine learning involves evaluating the project itself. For many businesses, getting through that step is a challenge.
Allen suggests a framework for getting started:
- Assess internal support: Is it low hanging fruit? Do you need a quick win? Are you just getting started? Can you measure the rewards? Everyone thinks everyone else is doing machine learning, but in reality, very few are.
- Figure out types of business impact: It reduces costs, automates jobs that shouldn’t have been done by humans in the first place, and achieves breakthroughs. How you plug machine learning into your business in a way that achieves a measurable, positive outcome?
- Define the business problem in ML terms: You must be able to translate the business question into a machine learning question. What’s the form that a data scientist could actually work on? Choosing your question is not a trivial concept either. It must be accurate and measurable, or your initiative will be way off.
- Consider your data readiness: You must consider your data like a yearly physical because it’s not a static concept. As data comes in, it changes over time, and a data scientist is going to be very sensitive to anomalies or problems with the data. If your organization isn’t thinking like a data scientist, there’s a good chance there will be gaps in your data.
- Augment your people, don’t replace: Yes, there will be some job loss, but again, these are jobs people shouldn’t have been doing in the first place. Leaning into the message is part of getting machine learning on the table in your organization. Your people have to understand how it will benefit, and not ultimately replace, their labor.
- Have realistic expectations for accuracy: It’s probabilistic. There are no definitive answers. Instead of being rules-based, it’s merely predictive based on patterns of behavior. Having 100% accuracy isn’t the goal; instead, having a higher percentage of accuracy is the goal. Ideally, you’re on the same percentage as what humans can do or a little better.
Getting to Deployment in the Business Context
The best projects have measurable business impacts and a well-defined question to get you there. Machine learning isn’t overhyped, but there must be some realistic expectations in place to make sure you’re undertaking a project in the best faith you can.
[Related Article: How Product Managers Learn About AI Meeting Peak Effectiveness]
According to Allen, the best ML projects will have:
- measurable impact
- data readiness
- team readiness
- well-defined question
- realistic expectations
Centralize your data function and make sure that you have the expertise to get the right kinds of data to apply to your questions. Machine learning is going to help us finally tackle problems that humans weren’t great at solving alone. It’s all about making sure you’re enterprise ready.