As a fan of the show Mad Men and its wonderful anachronisms, I had a good chuckle over Sterling Cooper & Partners’ season 6 acquisition of the IBM System 360. Without spoiling much for anyone who hasn’t seen it, the firm attempts to step up its data research game by bringing in a computer mainframe so mammoth it takes over the entire employee lounge, and seems so alien and imposing that one staffer worries it might actually be reading his thoughts.
The introduction of the IBM 360 50 years ago was actually revolutionary, and it helped pave the way for the modern computing systems we use today. And I’m sure that generations of ad men believed it was an effective tool for winning accounts and selling more soap. But in terms of capabilities, it probably did less than a first-generation iPod shuffle.
Building a Better Algorithm
In another 50 years, today’s computers will no doubt look just as primitive and limited as the IBM 360. Whatever leaps made between now and then will likely be due to something known as machine learning. Machine learning is basically the difference between a computing system that simply stores and retrieves data and one that can recognize, group and sort data without an explicit program.
Some people think of machine learning as a form of artificial intelligence—algorithms that are driven by data rather than programming, which can improve and even evolve over time without human intervention. Machine learning is a big part of what makes smartphones smart, of how search engines are optimized to our tastes, and of how we can trust that our spam folders will probably be filled with spam. You know the feeling that Facebook, Pandora or Amazon know far too much about you? That ability to predict what you might want to see before you tell it to? That’s machine learning.
Supervised and Unsupervised Machine Learning
As smart as our computers are, they are still only as good as the human-generated algorithms that instruct them. The ways the algorithms “learn” generally fall into two categories: supervised learning and unsupervised learning.
In supervised learning algorithms, the data is classified or labeled in a way that helps the program make decisions. Unsupervised learning algorithms, on the other hand, make connections without input or labels. Unsupervised learning involves the algorithm clustering data to discover similarities on its own. If a program includes a combination of both, relying on classifications as well as clusters, it may be called semi-supervised or active learning.
To help make sense of the difference, I found a good, easy-to-understand example on the boards at Stack Overflow. Say the data consisted of photographs. A supervised learning program would classify what is a face and what is not, and the algorithm would eventually learn the difference. An unsupervised learning algorithm gets no such classification. The algorithm itself must cluster or group different types of data: faces, landscapes, and so on.
Big Data Wants You
There are people who would balk at the term machine learning, who would say it simply does not exist. Machines don’t learn, and as intelligent as a computer might seem, it is all algorithmic illusion. Machines are only as smart as the human beings who program them. All of which is true. But it’s still pretty incredible what algorithms can do.
Whether or not you agree with the term or the theory, machine learning looks pretty good on a resume. Experts in this field are in such high demand, one McKinsey study estimated that need for people who understand machine learning could be up to 60 percent greater than supply by 2018.