Machine Learning

Is a robot uprising taking place?

Is a robot uprising taking place?

The short answer is no.

The long answer is…

They’re working on it. Machines learn to act a certain way by application of what’s called a cost function. The cost function is integral in a machine learning how to solve a given problem. What the cost function does is when a machine tries a certain action, the cost function will reward or penalize the machine for the action taken.

Machine Learning and Cost Functions Explained

A simple example of this might be if a machine was learning to play a video game. If the machine made an action that lead to it getting more points, you would reward it 1 point. If the machine made an action that lead to it dying, you would remove a point.

Machines Will Try To Cheat When Allowed

Sounds simple right? This is where science fiction kicks in. All those movies that decide to eliminate mankind were on to something. For instance, in my Coursera course there is a famous example of an AI algorithm that evaluated if a CT scan contained a benign or malignant tumor. After “training” it with a cost function similar to the one described above, it came up to be 99% accurate. Impressive right?

As it turns out, the CT scans that it was evaluating contained 99% benign tumors, which means the algorithm always guessed benign, making it right 99% of the time.

When we train a machine to just get the best answer, it will find the shortest path to get to that correct answer. Sound like another species you’ve heard of? This means that we have to be very careful about how we reward our machines for their answers.

Google Photo Tagging Faux Pas

Another great example is when Google’s photo tagging algorithm tagged two black people as gorillas. The way this algorithm was trained was through looking at other pictures that were tagged. My best guess is behind the scenes, it give a probability that the picture was gorillas or people, and gorillas were highest by a tiny percentage. The algorithm was designed to choose whatever is a higher percentage regardless, so it did. There was no sensitivity to this choice built in, so it acted logically.

An improvement to this might be for the tagging algorithm to understand that if there is a possibility the picture could be people, to not tag the picture at all. Teaching a machine what pictures it should be most “certain” of before tagging is just part of the learning.

What’s Being Done About This

There are professors now writing lots of articles about writing BETTER cost functions. One article I read was about supervising the machines as they learn in a more proactive way, so when they start to cheat in creative ways (because they will) it can be caught quickly and changed.

There is also studies happening watching what happens when we allow machines to cheat in certain circumstances. In one example, several linguistic problems were tackled with two different cost functions. The first cost function was VERY strict about what it rewarded and kept cheating to a minimum. The other was more loose in it’s rules. Across different problems, it was found more often the one with stricter rules found more creative ways to get the correct answer, but also sometimes wouldn’t solve the problem at all. Depending on the problem, a creative, innovative solution might be favorable to have a machine get to.

Yet another example is a strategy called adversarial learning. Basically, two machines compete. One is set to solve a problem, and the other is set to watch the machine solving the problem and catch any incorrect responses. This strategy would be a great way to catch cheating algorithms.

Breath: We’re Going To Be OK

The long answer is that the people doing the bleeding edge research have identified the problems of machines trying to cheat and work outside their solutions. In business, it’s also more profitable to have really good cost functions to have your machines optimizing to the solution you want. With both science and business aligned in creating better cost functions, I feel safe knowing that this field will be given more attention, and therefore, with better cost functions, comes machines that behave as we want them to.

Robot apocalypse averted.

Disagree? Leave a comment below!

Have I misrepresented something? Let me know in the comments!

Leave a Reply

Your email address will not be published. Required fields are marked *