We invited Ambica Rajagopal, Group Chief AI Officer, Michelin to share her thoughts on “Interconnecting Data Science With Business To Benefit The Enterprise" at our recent Data ROI event. You can watch her full session here or read on for a few highlights.
Explainability is Key To Generating Buy-In
Fear of the unknown is intrinsic to the human experience. While applications of artificial intelligence (AI) are now commonplace, few people really understand how AI works.
Though often associated with AI ethics, AI explainability can also be a powerful tool for generating buy-in from the business. As defined by IBM, explainable AI is “ a set of processes and methods that allows human users to comprehend and trust the results and output created by machine learning algorithms. Explainable AI is used to describe an AI model, its expected impact and potential biases.”
In addition to driving trust and confidence when it comes to outputs, explainable AI helps non-technical employees truly understand how AI can help deliver tangible business benefits. Explainable AI methods not only ensure people understand how AI models come to their decisions, but also shed light on how these outcomes should be interpreted.
As Ambica explains, “There are aspects of AI, that even from a technical perspective, are not easily explainable and have a black box nature. So getting the organization or the folks who are making decisions, using this technology, to be able to trust it is critical for long-term success.”
She adds that explainability is built into the DNA of Michelin’s AI practice. “We even have guilds that we have established around specific technologies. These are communities of practice around natural language processing, computer vision, and so on. These guilds include folks who are not data scientists, but business professionals and this is where both sides of the AI equation can work together and share information.”
Invest in AI and Data Cleansing at the Same Time
One mistake many AI creators make is that they invest heavily in overhauling their data before they really, truly understand what they want to do with it. Ambica tells us, “until you've taken apart your data through the lens of usage, you don't really know whether it's of the right quality or whether it's really what we want. In fact, are you missing some of the data that you need? Are you just measuring it wrong?
And so embarking on a large data cleansing effort, with very minimal investment on the AI side, can lead to such situations where later on down the line, you find yourself at a place where you have to in fact go back and take a look at some of your data collection methods and such.”
She later adds, “expecting an organization to be sort of data ready, before embarking on an AI journey, might not be necessary in fact. It's something that could actually be thought about together.”
The Point of AI is to Augment Human Behaviors, Not Replace Them
There’s no doubt about it: AI will redefine how we work and, yes, it will displace some workers. However, more often than not, AI will merely enhance the employee experience, not take over it. In fact, the term “artificial intelligence” is rather misleading as AI models don’t “think” but calculate.
“At Michelin, our AI program is intended to augment human capabilities and intelligence, and the presence of a human layer of decision making on top of our AI engines is something that we view as critical to success,” Ambica explains. “If you look at the technology overall, I think there are two aspects to it. So one is an aspect of automation. And this certainly, we have quite a bit of it, but at the same time, there is a heavy effort that goes into making sure that where we have automation through AI, there is in fact a good layer of manual check, whether it is from a labeling effort, whether it is from that is incorporated into the MLOps parts of our, let's say a CVL, lot of them in production.”
Something we don’t think about enough is that AI algorithms are universal approximators, they're able to model processes in multiple dimensions. This is something fixed models, physical models have not been able to do in the past. This is something human beings are not very good at - finding correlations, honestly, even within two variables. That's one of our blind spots and AI models are able to give us insights into processes that otherwise would not have been obvious to us.”