The exciting “rocket science” part of any predictive AI project may be training the model, but without the proper focus on deployment, the rocket will quickly stall.
At NRF Nexus, Eric Siegel, founder and CEO of Gooder AI, said machine learning projects that aim to improve business operations by generating predictions often fall short of their intended goals.
Instead, we fall into a “no man’s land” between business professionals and data professionals, each claiming that it’s not their role.
Data scientists believe the value of the models is obvious, so they'll definitely be deployed, and anything else might be considered an “administration issue,” he said. Meanwhile, business stakeholders are saying, “No. That semi-technical level of detail? We'll outsource all that stuff. That's what we have data scientists for.”
That's why “hoses and faucets regularly fail to connect,” Siegel said. “It's a real irony: this emphasis on rocket science at its core has us more excited about rocket science than we are about rocket launches.”
But from a business perspective, deployment is the big deal, it's everything.
“Operationalization, productization, deployment in the field, change in operations — that's where the benefits come from,” he said.
“Number crunching alone doesn't give you value. Machine learning AI isn't inherently valuable. It only becomes valuable when you act on it, and that's deployment – actually making operational changes based on predictions to improve operations.”
Siegel offered an alternative approach: BizML, a set of business practices designed to enhance collaboration and drive success in applied machine learning projects.
Of the six steps, the last three are “universal” – prepare the data, train the model, and deploy the model – but before that there are three more “pre-production” steps: set the deployment goal (value proposition), set the prediction goal (more detailed), and set the metrics (both business and technical).
Gooder AI is currently in early stage product trials using an interactive interface to translate technical metrics into KPIs, allowing users to visualize and explore possibilities.
He advises business people to upskill on the semi-technical aspects so that they can be active participants in the process from start to finish. He says that too often machine learning projects focus on technical metrics like precision, recall, accuracy, etc., instead of business metrics that everyone can understand, like profits, savings, and ROI.
AI Working Group
This retail member-only working group will provide feedback on proposed AI laws and regulations, advise NRF on developing common principles and guidelines for the use of AI in retail, and help NRF educate policymakers.
learn more.
“People in the business will also say, 'You don't need to look under the hood to drive a car,'” he says. “And that's true. I personally don't know much about engines. I don't know where the spark plugs are. But when it comes to driving, I'm a total expert.
“I know the rules of the road, the momentum, the friction, how to operate a car, and what drivers expect from each other. And now, that same level of expertise is absolutely necessary for any machine learning project to be successful.”
For those daunted by a semi-technical understanding, Siegel offers a more detailed explanation in his recent book, “The AI Playbook,” which he calls “much more accessible, generally applicable, interesting, exciting and understandable than high-school algebra. It's just not yet part of the co-curricular curriculum.”
Siegel also presented case studies from UPS and FICO, comparing generative and predictive AI. While generative AI may be easier to use than predictive AI, Siegel said it's “harder to use well.”
