What Machine Intelligence Can Do

Part 2 of "Separating Hype from Reality," Chapter 2 of The Machine Intelligence Primer

Let’s start with the good news: MI research over the past 60 years—and especially the past 5—has given us some truly incredible machine capabilities.  

Since the early days of MI, researchers have developed techniques for enabling machines to perform simple, rules-based tasks, usually by codifying human knowledge as sets of logic-based rules and programming rules into computers (“symbolist machine learning”). Expert systems are the quintessential example of this early MI technology. Using the knowledge of accountants, doctors, and other types of human experts, researchers built machines that could answer straightforward questions with precision and scale. Many commercial applications of these systems went out of fashion after the 1980s, due in large part to their rigidity and labor intensiveness. Collecting and formalizing the knowledge of an expert takes a great deal of time, and each time the experts’ knowledge evolves, the machines require updates. Useful applications of expert system-like products do, however, persist. For example, tax preparation and accounting software can be viewed as an expert system.¹

New applications of early machine learning research that rely on the same, basic concepts are also still emerging. Robotic Process Automation (RPA), which has grown increas- ingly popular over the past 3 years, is one such example. In RPA, virtual software “bots” are programmed onto machines to “observe” clerical staff performing simple, high-volume, rules-based tasks. Through observation of their human counterparts’ digital behavior, the bots develop rules they can use to repeat the steps themselves, autonomously. RPA is preferable to more classic computer programming-based automation because it adapts to specific orga- nizations’ unique processes. RPA is also easier to update than computer programming, as bots can be reconfigured to retrain when an organization’s processes change. Organizations are adopting RPA quickly, seizing an opportunity to perform rote tasks more accurately and quickly and enabling staff to spend time on more complex, value-added activities.

While simple task automation has great potential to save organizations time and money, the current excitement surounding MI is largely the result of significant improvements in MI’s ability to find and act on patterns in data (“Pattern Recognition”). In the past 5 years, enabling technology trends have made available massive amounts of data and computing power with which to train machine learning algorithms. The development of sophisticated deep learning algorithms inspired by the circuitry of the human brain has further boosted MI’s ability to find structure in data.

The math and computer science behind the developments that enable machines to find patterns in data is complex, but the applications we have today are relatively straightforward. Machine learning algorithms ingest, learn from, abstract, and act on data in an “Input Data A generates some Simple Response B” fashion.² Most of the tasks pattern recognition MI executes could be completed by a human with a few seconds of thought.

Given the relative simplicity of pattern recognition MI, it may be surprising that it generates such technological euphoria. But, a larger portion of tasks that we perform than one might expect take just a few seconds of thought. Strung together, these moments-long actions make up a meaningful portion of our personal and working lives. A doctor reviewing an x-ray of an arm takes just a moment to determine if there’s a break. Driving down the highway, we know almost instantaneously whether we have time to merge lanes. A fluent translator takes just a moment to convey what was said by one party in English to a colleague in French. The applications of pattern recognition data are expansive (Fig. 2.3), and for many industries, transformative.

An accounting of the many applications of pattern recognition MI makes clear why it is inspiring such enthusiasm and investment on the part of government and private sector organizations. It also elucidates growing concerns over how MI will impact our jobs.

It’s important to note, on this front, that while many jobs—from those in radiology to trucking—will be significantly impacted by MI, relatively few will be eliminated entirely. Simple, straightforward tasks are building blocks of many of our jobs, but they don’t form the full picture of our working lives. They fit into a broader context of more complex, higher-order cognitive tasks and human interactions—and today’s machines aren’t very good at context.