Putting AI at work
Most of the challenges companies face around AI are really around their data
Image: ShutterstockQ. In addition to teaching at the Rotman School, you work for a highly successful start-up. What problem is your company, Rubikloud Technology, trying to solve?
Rubikloud is a venture-backed start-up company in Toronto that uses AI to enable intelligent decision automation for some of the world’s largest retailers. Our goal is to build AI software systems that can take the hundreds of decisions that a retailer has to make every day and automate and optimize those decisions. We’ve been working with some of the largest retailers around the world, from Asia to Europe to North America. Our most recent venture round, led by Intel Capital, led to a total investment of $45 million, which we hope will enable us to change the technology landscape for enterprise retailers.
Q. Can you describe what exactly your product does?
There are two main functions. First, on the customer lifecycle side, it provides an understanding of the customer on an individual level for targeted promotions and marketing activities that encourage them to shop at a particular store. Second, it addresses the mass market and merchandising side, where we help retailers forecast and optimize the different promotions they do in-store. When you are at the grocery store and you see a certain brand of milk on sale for $2.99, there is actually a complex process in the background to make that happen. We help retailers select the right products to put on promotion, determine what those promotions should be as well as forecasting to make sure they have the right inventory at the stores.
Q. As Chief Data Scientist, what is your role in all of this?
I’m responsible for all the data science functions, which primarily revolve around building out our machine learning and artificial intelligence capabilities. We have two key areas of focus: applied R&D and fundamental R&D. The applied team is responsible for researching and building out the core of our AI system in the kind of products I mentioned above, which includes machine learning models and the underlying software systems that allow us to automate, scale and store all the data and models.
The fundamental research side is more focused on strategic long-term issues for the company. The things they work on may not even appear in our products in the next five years, but we recognize that we need to look ahead and have a long-term roadmap as to what our products will be able to do one day. We’ve also begun to engage with the academic community, and are hoping to publish some of our research in the near future.
Q. Tell us a bit about your career path, and what led you to this role.
I joined Rubikloud early, in the first year of its existence. There’s a popular saying, that ‘one year at a start-up equals three years at a bigger company’. I’ve been here four years, so maybe that really means 12 years? But seriously, one thing that really helps to accelerate a career is being involved at an early stage with a high-growth start-up. There are opportunities to grow and learn that frankly, you can’t find anywhere else. I have basically watched this company grow from a half a dozen people to over 100 in a very short time span. We’ve been doubling our headcount pretty much every year. In that time, I’ve done some engineering work and even some client facing stuff in addition to my data science work.
The key thing about hyper-growth start-ups is that if you’re going to thrive in one, you really have to be quick to adapt and learn. In terms of management, the approach that worked with 25 people will need to change when you get to 50 people, and that approach will not work when you get to 100 people. You really have to pay attention, not just to the technology, but to the organizational and management aspects. I’ve been very fortunate along the way, because I’ve had great mentorship from our CTO, Waleed Ayoub, whom I report to—and who gave me the opportunity to step up and run the data science side of the business.
Q. What does it take for an organization to make the most of its data? Are there some principles that you can share?
It actually takes a lot. The elephant in the room—especially with respect to AI—is that it is impossible to have good AI without good data; and good data is a really hard thing to achieve. From collecting it, to cleaning it, to building the right technology platform so that you can access it in a scalable way—it’s very involved. To give you some context, for a large enterprise to modernize its data practice, it could take tens of millions of dollars and involve dozens of people over a multi-year time span.
In my experience, most of the challenges companies face around AI are really around their data. The first step for leaders is to understand and accept the complexity of getting good data, and start treating it seriously. The focus on this should be on the same level as financial reporting and results. If you are going to become a data-driven company, it takes a lot of work.
Q. Are businesses embracing AI and machine learning to the extent that they should?
I work with enterprise retailers, but what I see with them is probably typical of most companies: They are approaching this new technology carefully and with measure. The hard part is that they are getting squeezed from both ends. A few really big players are dominating the industry, like Amazon; then, there are all of these smaller E-commerce start-ups targeting niches that are also gaining a lot of traction—and both sets of challengers are using AI. This has really opened them up to making the transition.
In many cases, it’s difficult for an organization to do these things internally. Changing existing business processes and legacy systems and hiring the right people is very difficult. That’s why a lot of companies are willing to partner with smaller companies to help them make the transition to AI and the next generation of technology.
Q. Tell us how you became involved at the Rotman School.
A good friend of mine introduced me to Mihnea [Moldoveanu], the Vice Dean of Innovation and Learning at Rotman. We both have an engineering background, so we got to chatting. He was especially interested in the AI work that I’d been doing, and he identified a couple of related initiatives at Rotman: the new Master of Management Analytics Program and the newly created Management Data Lab. He told me they could really use someone from the industry who has AI expertise. From there, I met with Dmitry Krass, the academic director of the MMA, as well as Susan Christoffersen, Vice Dean of Undergraduate and Specialized Programs. We all agreed that it would be great for students to have someone with industry experience helping out with the program, as well as in the data lab.
I started working with them last April. Everyone is really passionate about trying to make the MMA Program as good as it can be to help students bridge the gap between technical expertise and the management side, as well as bringing the School’s data science practice into the next generation by building up the Management Data Lab. On the MMA, side I’ll be teaching a couple of modules on deep learning and neural networks, which have been all the rage in the past little while. And I’m also working with the Management Data Lab to try to understand what the needs are within Rotman, and helping build out the hardware and software capabilities to support data science teaching and research at Rotman.
Q. Some people argue that data can’t solve problems because by its very nature, it is information about the past, not the future. How do you react to that view?
I’ve heard that before, and I agree with it to some extent. It’s kind of like what Yogi Berra said many years ago: ‘It’s tough to make predictions, especially about the future’. What is missing from this perspective is the idea that data can be used in many different ways—and for a certain category of problems, the future actually does resemble the past. Obviously in such cases, data is applicable.
However, I would argue that even when the future is more uncertain, data can help a problem solver quantify that uncertainty. For example, it can’t tell you, ‘it’s definitely going to rain tomorrow’, but it can tell you ‘there’s a 70% chance of rain tomorrow’. In the realm of business, it’s not able to say ‘this customer is definitely going to buy this product’—but it can say, ‘there is a high chance that this customer will buy something’. Even if you’re not sure exactly what’s going to happen, it can inform you about something that otherwise would be left to intuition.
I do agree that we shouldn’t blindly use data to solve all our problems. We need to understand the ways in which data is useful and use it judiciously. It’s not like we’re ever going to throw away human judgment just because we have data; but the fact is, data can be a huge boost to many of the decisions we are making.
Some people still fear that machines will take over most of our jobs, but in my view, they needn’t worry. Take medical diagnosis as an example. Already, machine learning algorithms do a pretty good job at diagnosing—and in some cases, they do even better than human doctors. But my wife is a physician, and she understands that diagnosing a patient is not just a mechanical set of decisions. It involves a lot of human interaction and observation, which a machine can never replace.
Furthermore, the technology is nowhere near being able to provide ‘artificial general intelligence’—whereby a machine could potentially replace a human mind. And as indicated, many organizations are limited as to what they can accomplish, given the data they have. A machine is only as good as the data used to train it, and we have a long way to go on that front. There are also some things that we just can’t collect data for, or that it’s too hard to collect data for, and as a result, at the end of the day, a machine is never going to have the judgment that a human has. Having said that, humans do also have their own biases and other deficiencies, so I believe the strongest approach is a marriage between human and machine capabilities.
Q. If you could change something about the current entrepreneurial landscape, what would it be?
Especially in Toronto, I would have a lot more Silicon Valley-type venture funding. The culture and calculus of how Canadian investors think is fundamentally different from their Silicon Valley counterparts. I’m generalizing a bit here, but in Canada it is much more risk-averse. Investors tend to be looking for a ‘high likelihood investment’, whereas in the Valley, they are very much focused on finding ‘the next big thing’—the next Facebook or Uber—and that leads them to invest in more risky propositions. They truly do not care if 19 out of 20 fail, as long as one results in a big win. That’s the mindset we need to have if we’re ever going to generate the next big tech start-up in Canada.
Brian Keng, PhD is an Adjunct Professor of Data Science at the Rotman School of Management and the Chief Data Scientist at Rubikloud, a venture-backed AI startup in Toronto. He teaches in the Rotman School’s new Master of Management Analytics Program.
[This article has been reprinted, with permission, from Rotman Management, the magazine of the University of Toronto's Rotman School of Management]
