Understanding Understanding: Why Machine Learning Can’t Work Without Humans



Much like old dogs, it’s hard to teach machines new tricks — but that’s because machines don’t learn.

At Media6Degrees (m6d), we build scalable systems to target customer prospects for major brands.  But contrary to convention in the industry, I refuse to call those systems “machine learning.”  The machines are not learning – or at least they’re not “understanding” in any sense comparable to human understanding.  These systems are using predictive modeling in its barest form.  We use data to approximate a function that maps a person’s partial browsing history into a brand signal, and that signal gauges how likely the person is to buy a product (or to take some other desired action).  While the systems are technically full of challenges, they are not “understanding” the people they evaluate.

But to be fair, the expectation of understanding is really a very tall order.

The challenges of understanding are illustrated nicely comparing the meanings of two nearly identical sentences:

“Rob was mad at Tom because he stole his lunch”

vs.

“Rob was mad at Tom, so he stole his lunch.”

Even our smartest computers can’t identify with any confidence who is ‘he’ and who is ‘his’ in each of these sentences.  But the truth is that language is a people invention, so it’s no wonder computers have to play catch-up.

On the other hand, people are terrible at understanding probabilities.  It is just not in our DNA to relate to uncertainty.  By contrast, computers are masters at understanding uncertainty!  They can tell the difference between a person who has a 2.9 percent probability of buying running shoes in the next month vs. a person with a 1.3 percent probability.  That is a valuable and important distinction for a marketer, but it is generally invisible to the human brain.

Let’s face it — the complex math involving models, formulae and algorithms are “skills” that are easy for computers, but difficult for the majority of average humans.  And while marketing professionals can be either right-brained or left-brained, depending on their role and responsibilities, very few marketing professionals are also advanced mathematicians.  But then very few computers (if any) can “surprise and delight” with their dazzling designs and brilliant copy.

So the truth is, in marketing, computers and humans have to rely on each other.  We can’t expect computers to understand the finer points of our complex language any more than we expect them to understand the emotional connection we feel to a Nike ad.  By contrast, it would take a human several hours and many sheets of paper (and, presumably, many erasers) to determine how much that emotional Nike ad raised the probability that a particular consumer will go out and buy their shoes. The computer will understand that a probability existed, that an event occurred and increased that probability, and that the probability was increased by an increment.  It will not understand why LeBron James was featured in the latest Nike ad, or why their cricket ads aren’t displayed in the U.S.

You still need humans for that.

Claudia Perlich is an award-winning data scientist who joined m6d as Chief Scientist in February 2010. In this role, she designs, develops, analyzes and optimizes the machine learning systems that help brands find their best prospective customers. An active industry speaker and frequent contributor to industry publications, she thrives on serving as a guide in the world of data. She has been published in over 30 scientific journals, and holds multiple patents in machine learning. P She has a PhD in Information Systems from NYU and an MA in Computer Science from the University of Colorado.

•