Every prediction of the future has artificial intelligence. But although computers have beaten humans at chess, Go, and more, we still claim that artificial intelligence is something that does not yet exist. Instead, what we have is "machine learning."

Computers are tools for following rules very quickly. Typically, humans create these rules. Machine learning is where the computers figure out the rules for themselves. To see how this works, you will first teach a computer how to sort these colored balls, based on the brightness picked up by a camera and the size of the ball. Try to select all the small, green balls.

(Try to get the small green balls on the left of the divider, and the other balls on the right)

This is doable. But real world tasks are far more difficult. Let's say you're a bank, and you want to find out which mortgages to approve. Can you teach the computer how to figure out which ones to approve based on statistics about the applicant, such as their income and the income of their neighborhood?

(Try to get the end of the bars corresponding to approved mortgages on the right of the divider, and the end of the other bars on the left)

That was a little too hard, right? Although we know these are all factors, we don't fully understand how they combine. But computers are very good at doing things fast. This is where machine learning comes in. If we show the computer hundreds of examples, we can adjust the knobs every time it gets one wrong, based on how much that knob contributed to that mistake. This will automatically correct for errors over and over until the predictions are extremely good.

But what exactly did the computer teach itself? Since we didn't tell it what to do, in order to understand why it does any particular action we have to guess based on what it does to similar inputs. For example, let's take a look at what percentages of mortgages are approved by our algorithm for these two groups:

The left group has around twice as many mortgages approved! What's the difference between these two groups?

The left group is all the white applicants, and the right group is all the Black or African American applicants. Did we just create a racist AI?

Apparently, but we didn't include race anywhere. How did this happen?

The answer is that it's hidden in our "correct" answers that the computer trained itself from. Because those were based on decisions made by people, who have society's ingrained biases, the bias is hidden in the machine learning model that the computer created. The algorithm is not neutral.

How do we fix this problem? If you can figure that out, please let us know. But be aware that machine learning models are used in situations including determining sentencing of inmates, computing housing prices, and determining what ads you're shown, even political ads. For more information on this, read this.

In real life, hundreds of tiny models like this are combined in ways that let them learn more and more complicated things about the data, which are combined to be able to take things into account in relatively smart ways. To learn more about this technique, called deep neural networks, read this, or for a slightly more technical explanation, read this.

Machine learning can also be used to create things, such as images (some examples include DeepDream, FaceApp, neural style transfer," and the recent crisis involving fake porn of celebrities) or text (for example, Google Translate). To learn about how this works, read this (this article does assume some familiarity with coding).

HOME