Working Up

by Dwayne Phillips

Despite all the promise and all the already fulfilled promise, there is a fundamental problem with machine learning.

Machine learning is that part of artificial intelligence that time and technology favors. We have the stuff needed to make machine learning work right now.

Machine learning “works” because you feed in a million photos of things (at least for computer vision, the other ML areas have the same situation), learn, i.e., create a model, and then use it.

The technique can “read” x-rays, find cracks in airplane engines, warn folks of impending forest fires and all sorts of things that benefit people.

Then there are those applications where the technique finds a criminal in a crowded room by recognizing a face or points to someone who “doesn’t belong.” We find that there are all sorts of “biases” in those million photos that some of us don’t like in some situations. What do we do?

Solution: go through those million photos by hand to ensure we have a set that is “fair” or “better” in some way. Going through all those photos by hand defeats the purpose of the entire exercise. We will be inspecting everything manually to ensure that the machine does everything automatically. Huh? Won’t work.

Well, we can still justify this as you have to look at a million photos once to create a good “model.” Afterwards, you use the model over and over and over. That reuse has a greater return on investment than having people do all the work from now on. Really? Have you run the numbers on that? Well, uh, …

There is something some of us call “the Technology Imperative.” It states that if the technology and technique are available, we must use them. That usually wins the day.

Perhaps something good will come of this.