How to Apply and Optimize Your Algorithm When You're Ready to Run With AI
Amazon’s recently launched SageMaker artificial intelligence service is an exciting new development, but the program doesn’t do it all. There’s a distinct gap between innovative AI technology that exists and AI solutions that will help drive business results in your specific case. Using products such as SageMaker is like having a brand-new Tesla Model S: It’s an awesome car, but it’s a giant electric paperweight if you don’t know how to drive.
We discussed "walking" with AI in a prior Entrepreneur article; now it’s time to hit the ground running. At Manifold, we work with clients using a method called “Lean AI.” Our method is inspired by many other popular processes, including human-centered design by IDEO, agile software development, the Lean Startup methodology and CRISP-DM. Lean AI has six steps: understand, engineer, model, acquire feedback, deploy and validate. Here, I’ll focus on three key pieces that any entrepreneur will need to follow to optimize AI.
1. Engineer: Quit playing around
Because AI engineering is software engineering, you need to use good practices such as source control, code reviews and clean interfaces, among others. Many data scientists are guilty of “playing in the sandbox,” but you should always build as if you’re going to production.
At Manifold, one of the most important steps we’ve implemented involves using Docker to take advantage of containerized data science. The resulting developer flow is cleaner and more collaborative, and it's ultimately far more productive.
People have been engineering software far longer than they’ve been engineering AI solutions. Applying existing development and operations best practices to AI systems will make your processes as efficient as possible.
2. Model: Start small, and scale up
When incorporating AI into your business, the possibilities seem endless. Don’t let your imagination get the best of you -- even if you have big plans, you’ll want to start simple and scale up. Take the advice of Emmanuel Ameisen, the AI lead at Insight Data Science, a post-doctoral fellowship program connected with big Silicon Valley names like Facebook and Zillow: Efficient problem-solving happens at the most straightforward, basic level. Baseline models will consistently deliver superior end products, especially for the user.
We use explicit rules in our process to keep simplicity in mind, particularly when applied to supervised learning problems. We believe in nailing a few features first; you can always add more later. And we always begin with classification before regression -- dealing with a set number of values rather than a continuous values -- so we can learn from the more obvious class errors.
3. Acquire feedback: See what (many) humans think
At the end of the day, humans will have to interact with and make sense of your AI’s recommendations. Get your AI in front of users -- fast. They can tell when the AI is recommending reasonable things or the search results are relevant. In our research, we’ve found two major patterns: suspicion of AI and the need to post-process raw predictions.
AI models rarely gain immediate trust, especially among people who haven’t worked with machine learning before. Explainable AI is still a field in its infancy, but there are great packages already, like Tree SHAP, that explain the “whys” of an AI’s predictions so users feel more comfortable.
We’ve also found that an AI’s raw predictions are often insufficient on their own. It’s necessary to build a user interface that allows post-processing so users can go a little further to solve the business problem. One of our clients, a leading oil field services company, had many compressor units that were running in "stressed" situations. While raw AI predictions correctly predicted that these units were going to fail soon, that wasn't useful information to the maintenance techs. Rather, they wanted to find "normal" units that transitioned to "likely to fail."
To solve the problem, we post-processed the time series of raw predictions and built a user interface that only alerted techs when a unit had significant changes in its failure probability. This resulted in fewer false alarms and a more useful AI with fewer inefficiencies. The takeaway for us has been that the user interface, or UI, is as important as the AI.
AI can accelerate businesses to new levels of insight more quickly than we know. However, AI and machine learning are still in their relative infancy. Because of the newness, business owners and entrepreneurs might be intimidated by the technology, or they might try to run with it before they can walk or even crawl. Instead, follow a tried-and-true implementation methodology of purposeful, simple and tested engineering, such as Lean AI, to unlock AI’s true potential.