When Machine Intelligence Meets Main Street

For data scientists and machine-learning experts, March 2016 was a momentous month. AlphaGo, a computer program developed by Google, beat world champion Lee Sedol at the ancient Chinese board game Go by a score of 4 to 1. In contrast to chess, where players might make about 40 moves per game, games of Go may have 200 moves.1 Whereas IBM’s Deep Blue was used to defeat chess grand master Garry Kasparov in 1997, computer scientists can’t calculate all the moves required to win at Go. Instead, Google had to create another kind of machine algorithm that could approximate humanlike qualities, playing the game by intuition and feel.

Significantly, programmers weren’t able to explain why and how AlphaGo made a certain move. Choices can’t be traced to the program’s source code any more than conscious decisions can be linked to a group of neural cells in our brains. AlphaGo’s latest triumph has therefore made clear that the rise of machines capable of self-learning is no longer just hypothetical.

We are past the point of debating whether human intuition can be replicated. Machine learning is already here. It will impact most companies over the next few decades and become part of everyday business life. Executives, regardless of which industries they are in, must quickly come to grip with how companies and industries will evolve. What should every manager know in the age of machine learning?

The Rise of Machine Intelligence

The idea of a thinking machine goes back at least as far as 1950, when British computer scientist Alan Turing wrote that if a machine was indistinguishable from a human during text-based conversations, then it was “thinking.”2 Computers, however, require programmers to write instructions. They don’t learn autonomously but follow rules.

The earliest iterations of so-called machine learning required heavy support and constant monitoring by computer scientists or statisticians. Data needed to be labeled, end goals explicitly set. In Amazon’s earliest days, it found that a machine algorithm called Amabot was more effective in generating customer recommendations to increase sales than humans who individually selected and promoted products. Amabot factored in customers’ previous purchases and their web searches.3

As capable as Amabot was, however, it couldn’t be used on different problems. Nor can algorithms be applied to unstructured data expressed in natural human language. The data needs to be in a relational database, such as an Excel spreadsheet.

AlphaGo can thus be seen as breaking the pattern of the past — the first machine capable of thinking, learning, and strategizing, all done with minimal human supervision. Before AlphaGo competed against a human, Google researchers worked on building a general-purpose algorithm to play video games such as Space Invaders, Breakout, and Pong. The machine learned the games by trial and error — pressing buttons randomly at first, then making the necessary adjustments to maximize rewards.

This approach to learning is built on a network of hardware and software that mimics the web of neurons in the human brain.4 With humans, reinforcement learning takes place when positive feedback triggers the neurotransmitter dopamine as a reward signal for the brain. Computers can be programmed to work in the same way, with numeric scores taking the place of dopamine.

This idea of “deep learning” is certainly not new — it has been discussed among computer scientists for more than 20 years. But recent advances in computing power have made it more feasible. It shouldn’t be a surprise that big tech companies such as Google, Facebook, IBM, and Microsoft have corporate labs dedicated to advancing machine learning. But what are the implications of machine learning for non-tech companies?

Linking Technology to People

Technology creators routinely bring forth new products and services that provide benefits to other sectors of the economy. Computers have eased office work; the Internet enabled e-commerce; airliners spurred tourism; automobiles encouraged consumers to shop at mega-stores; and containers facilitated global logistics. Although Sony, Canon, and Nikon don’t fabricate microprocessors themselves, they were able to commercialize digital cameras using technology developed by others. Viewed in this light, non-tech companies will have opportunities to embrace machine learning by making timely, strategic choices. Machine learning isn’t an end in and of itself; it’s an enabler. The key will be to find creative uses for it.

Malcolm Gladwell’s bestseller Blink celebrates the ability of a medical specialist to make a diagnosis almost as soon as a patient walks in or an art expert to “sense” when a work is a forgery. However, machine learning, as demonstrated by both AlphaGo and IBM Watson, is seriously challenging such dominance by human experts. For example, based on information from millions of pages of medical journals, IBM Watson can offer physicians recommendations, suggestions on additional blood tests, and updates on the latest clinical trials. To receive a diagnosis from Watson, all a cancer doctor with an iPad app has to do is describe the patient’s symptoms in plain English.

What happens when computing power, ubiquitous connectivity of sensors and mobile devices, and machine learning converge, equipping the world with a bevy of general-purpose, self-taught algorithms that coordinate our economic transactions? This scenario isn’t so farfetched, and the implications are far-reaching.

To begin with, transaction costs will shrink dramatically (if not go away entirely). When logistic coordination among industry players becomes automated, moreover, it is easy to see how redundancy in production facilities and waste will plummet, and the need for direct communication with managers will become less necessary. And once market coordination becomes easier, the argument for bringing activities in-house will weaken significantly: Whatever advantages large companies had for being vertically integrated will dissipate. This, in turn, will permit smaller players with far fewer resources to specialize in best-in-class services and deliver highly customized solutions to meet specific demands. Traditional companies must therefore prepare for a new economic reality that’s radically different from the one they’ve known.

Some important questions managers of non-tech companies should ask are:

1. Should I invest in machine learning in-house?

The paths to acquiring machine learning are limited: You can build, borrow, or buy it. Though it might be tempting to outsource machine learning entirely to third parties, this strategy effectively relegates the company to following existing market trends and only incorporating features that are commonly found in the marketplace.

Most companies, in contrast, seek to differentiate their offerings in the long run. Companies must therefore invest in an in-house team of data scientists who are able to repackage leading-edge technologies. They are the ones who will refashion emerging solutions made available by other technology creators.

2. How can I integrate machine learning into the company’s broader mission?

When pursuing a new capability, it is often difficult, if not impossible, for senior management to articulate a precise market strategy from day one. Experimentation and organizational learning are key. A bottom-up process works better than top-down. For this reason, companies should assign data scientists to business units as opposed to having them work in a centralized department. Only when data scientists are embedded as local advisors can the strategy making made possible by machine learning take full advantage of the changing opportunities of the local market environment.

3. How do I prioritize projects?

Given the great potential of machine learning, companies must reconcile near-term projects initiated by business units against investments with bigger long-term payoff. For most organizations, data scientists will always be scarce resources. However, even within that context, they will need room to experiment. Data scientists will likely prefer using open-source software to facilitate exchanges with external communities. But operating this way is anathema to patent lawyers and those charged with maintaining IT security, and it highlights why operational principles that seem innocuous can impede the development of machine learning. Therefore, executives need to play an active role in setting priorities and establishing the ground rules early.

The Road Ahead

AlphaGo may have been about games. But the broader implications are clear: We can expect imminent advances in commercial applications of many kinds. Demis Hassabis, who heads Google’s machine-learning team, has observed, “The methods we’ve used are general-purpose; our hope is that one day they could be extended to help us address some of society’s toughest and most pressing problems.”5

This is why opportunities for innovation will be abundant. Lest traditional, non-tech companies be left behind, they must cultivate a critical mass of individuals who are steeped in machine learning, place them close to businesses, and inspire them to do something bold.