A secret to adopting machine learning that has nothing to do with the actual technology.
It’s all about embracing an open mindset.
Machine learning has the potential to transform your business. To automate processes, uncover new insights, make your products and services better, and customers happier. Integrating the capability into your organization requires operational transformation and lots (and lots) of experimentation.
But, you know this already.
What you may not know is one key secret to making your initiative successful: keep an open mind. This is an expectation you must set early and reinforce often on the path to making machine learning operational for your business.
Getting started with machine learning
There’s a (somewhat scary) stat that we want to get out of the way: nearly 80% of all AI and machine learning projects stall out due to problems with data quality, labeling, and building trusted models according to a 2019 survey by Dimensional Research.
That hurts.
The remedy: to be successful with machine learning development, you must be incredibly intentional about the problems or opportunities you want to tackle. If you start out by solving something that is achievable and realistic, you can avoid being in this vast majority – and blitz your competitors.
To shape your goal, think on a granular level – what incremental, positive change can machine learning make to your processes or applications? What opportunities can you go after with this capability? When you have your list, consider every option with an open mind – don’t let the flashy, could-possibly-increase-revenue-by-89%-by-end-of-year idea completely distract you from moving forward with ideas that are more practical, at least for now. If you want to make machine learning operational across your organization, you need to start small. Build a model that you can test and iterate on without completely upending existing workflows.
So, what are some examples of good “starter problems/opportunities”? Examples vary by industry, but there are some general features that make early ML projects more likely to succeed:
- The application of ML should be assistive, ie. not trying to fully automate something off the bat. For instance, an algorithm to rank likely record matches in a record linking problem is a better start than trying to fully automate record linking – it keeps the human in the loop but speeds up the process.
- It should have a feedback mechanism so you can verify that it’s working after it’s deployed (are the top matches correct? how often is the correct result in the top 5?). Deploying any technology has a cost associated with it, so you’d like to know sooner rather than later whether or not it’s providing value.
Experiment, test, build, launch, fail, repeat
Once you’ve identified the opportunities you’re going to go after, it’s time to experiment.
You’ll need to build a multidisciplinary team with data scientists, SMEs, and those in charge of data governance. An SME can play a critical role in your team at the onset of experimentation. They can effectively help you navigate and access data (which is typically a big hurdle at the beginning of a project). Once data is collected, data custodians will build out information models and other data services to enable data scientists to be agile – building, testing, and iterating quickly.
At the same time, you must prioritize governance, security, and transparency – but in a way that still allows teams to explore and experiment. To keep the cost of governance down and the rate of experimentation high, consider a platform that facilitates continuous monitoring and security. This way your team can focus on innovation, opposed to the labor-intensive manual tracking and monitoring.
And remember: stay open-minded. With an open mind, the tough parts of making machine learning operational for your business don’t sting (as badly) and you may uncover opportunities you never saw coming in the face of failures.
Read Step 8 of our white paper, 10 Steps to Making Machine Learning Operational, to discover how your team may evolve once you find success with your machine learning models.
Editor's Choice
