When out talking to prospective enterprise customers we often find that the AI use cases that get the most attention in the technology press (e.g. chat-bots) fail to resonate with enterprises looking to leverage AI to solve more complex business problems. But it is not too hard to see how we got to this point. In current form, the easiest way to demonstrate the merits of cutting edge machine learning algorithms is by training systems to compete in games, solving “toy problems.”
Derrick Harris (@derrickharris) describes the key issue that arises from this dynamic in his most recent ArchiTECHt.io blog:
“I—and I assume a lot of CIO types—would like to know how reinforcement learning, for example, applies beyond defeating human champions at the game of Go. Particularly, it would be interesting to know how these types of algorithms might apply to some tough business problems. If you think hard enough you’ll conjure up some ideas but, frankly, most people don’t have the time to dwell on the subject, much less the expertise to start building anything.” (Why we play games with AI, Derrick Harris, ArchiTECHt.io)
Beyond games as a proving ground, the earliest actual applied AI use cases favor applications where there is a mountain of available data to leverage as the primary training source which helps explain the attention paid to virtual assistants, image recognition, and sentiment analysis. The reason for this can also be easily rationalized: in building these intelligent systems, the easiest way to train them is with loads of data. In fact, this pocket of AI use cases are a natural extension of big data, leveraging gigantic datasets to train machines to make faster and more accurate predictions and recommendations than we would otherwise make on our own.
But data is not the answer for every AI problem. Applications toward the left side of the spectrum in the graphic shown above are characterized by dynamic and expanding problem spaces (e.g. your Roomba doesn’t know your floor plan, the robot assisting in manufacturing of your car is retooled every model year, etc.). In these environments, more than raw data must be relied upon as the training sources if the goal is to build programmable, adaptive, and trusted intelligent systems.
There are at least two drivers of this reality. The first is money. The investment required to physically model and optimize the many different dimensions and variables within these complex settings quickly outruns the time, budget, and skill-set of many developers and enterprises. The second is knowledge. Enterprise systems and business processes are constructed, configured, and continually optimized in part based on input from an organization's most strategic asset: subject matter expertise. Despite what some would lead you to believe, this expertise doesn’t all sit neatly in some database or model. It sits in the spreadsheets and brains of the business analysts, the field engineers, the civil and mechanical engineers, etc. Capturing and codifying this expertise to create more intelligent and autonomous systems is not a big data problem. As such, for AI to realize its true potential, enterprises are going to have to think differently about solving these type of problems relative to the data-centric training blueprint that dominates the AI use cases covered in the media today. More concretely, this means that beyond throwing massive data sets against greater compute power and expecting systems to learn faster and more effectively, we must also come up with better ways to teach them.
In trying to understand how the enterprise embrace of AI evolves from here I am reminded of a comment made by James Hamilton, VP & Distinguished Engineer at AWS, during a presentation in 2011 regarding data center innovation. His basic point was that most great data center tech shows up in mobile phones first. But Hamilton’s insight likely fell short for those that couldn’t envision how a consumer electronics device could influence the roadmap of something as mission critical as the enterprise data center.
With AI innovation, it feels like we are at a similar point where games are the proxy, and favorite research proving ground, for the innovations that will eventually bleed into enterprise applications. The enterprises that will lead, and benefit the most from AI adoption will be the ones that can extrapolate real world use cases from the techniques demonstrated first within games. As an industry we must do our part to not only innovate around the AI platforms that will help bridge the gap between toy problems and business problems, but also educate enterprises as to the different lenses through which they can frame the problem.
For enterprises sorting through AI tools and technologies currently, the good news is that there are already very active pockets of innovation around AI development platforms that foster the construction of intelligent systems without relying solely on massive data sets. The use of simulations and digital twins, leveraging subject matter expertise, and reinforcement learning are just a few examples of emerging tools and techniques that will complement data in the rapidly evolving toolbox of the AI-enabled enterprise. At Bonsai, we are actively working with enterprises to help them leverage our AI development platform to build more intelligent systems and business processes. If you are interested in learning more about how your business can work with the Bonsai Platform to achieve this objective, please contact us to learn more.