Working Up

by Dwayne Phillips

“Machine Learning” is one of the biggest misuses of the English language I have seen. The machine isn’t learning. Let’s try to remember what is happening.

Data science is a big deal these days. Well, it is until you read the fine print in the Help Wanted ads about data science jobs. The fine print says AI/ML or artificial intelligence and machine learning. Employers want persons who can “build models” and know all the tool kits to do such.

And the term “artificial intelligence” has lost almost all of its meaning. It now means one thing: machine learning. We have machines that learn to solve all our problems.

Really? No, not really. The machine isn’t “learning” anything. The “model” isn’t a model of anything, either.

We have a set of equations. We add a few things together and see if the answer is “correct.” If it isn’t, we try again and again until the answer is correct. Each thing we add together has a number that multiplies it. Those are coefficients. The coefficients allow us to double one thing, halve another thing, take a third of another, and so one. Each time through our attempt at guessing the answer, we adjust these coefficients.

When we finally reach the correct answer, we take all these coefficients and store them on the computer disk. Those numbers we stored are the “model.” Huh? Why do we call it a “model” when it is just a bunch of numbers? Well, someone thought that was a good idea about 60 years ago.

Anyways, all this machine learning comes to to a situation where we create a loop in our logic. Each time we pass through this loop we multiply and add and see if we are correct. If not correct, we adjust all the coefficients, try again, and again, and again until the coefficients produce a correct answer. Viola’. We are there. By the way, the computer software to do this is quite simple.

At one time we did this with 10 numbers, then 100, then 1,000, and now with some arbitrary large number in the billions or trillions.

And we are more clever in how we do this than we were 60 years ago. And we have much, much, much better computers now that we did 60 years ago. Time produced computers that would work the solutions we created 60 years ago. Okay.

Still, let’s not kid ourselves. The machine isn’t learning. We are looping and adjusting coefficients until we have the result we want. That is pretty good stuff, but it isn’t “learning.”