Human Teachers Can Prevent Fatal AI Mistakes

Fagjun | Published 2017-09-06 17:41


 With artificial intelligence becoming more and more prevalent, people have been justifiably wary of getting replaced by a machine. Since AI is becoming capable of more and more things, there are fears that people are becoming obsolete, or that AI will overtake our capabilities in a few short years. However, it seems that these worries are unnecessary, as humans are still needed to fix AI mistakes.


Humans are still necessary to make AI safe and reliable.



Machines, artificial intelligence, neural networks—while these can be accurate and reliable, they can also make mistakes that can prove to be fatal. Even the best AI programs learn to be perfect by repeatedly making mistakes, but this kind of learning curve isn't ideal in certain cases. In AI programs for healthcare or driving, mistakes can be fatal.


Thus, there has to be a new way for AI to learn. Researchers have now figured out a way for AI to learn and perfect its program without serious failures.



Preventing AI Mistakes


Pong's inerface



A team of researchers at Oxford University have taken it upon themselves to train AI without it making a mistake. They had an AI program play Pong, a two-dimensional table tennis video game. To play the game, the player needs to control an in-game paddle and move it back and forth to hit the ball. You can either play against the computer or another person playing as a second player. The goal is to score 11 points before the other player does.


A human player would figure out right away that they would need to hit the ball with a paddle in order to increase the chances of scoring. An AI, however, would likely let the ball go past its paddle a few hundred times before it learns that it needs to hit the ball with the paddle.


Thus, a human has to teach an AI how to play the game. The researchers let another AI watch as a human player taught the other AI how to play. After four and a half hours of observation, the second AI was able to mimic the human player and prevent the first AI from making mistakes in Pong. Thus, it's possible to avoid AI mistakes while teaching it to do a certain task.



Safety Guarantees


Even human guidance won't make AI proficient in more complicated games.



These findings show that AI isn't actually on track to cut humans out of certain processes. In fact, humans are necessary for making sure that AI programs don't make mistakes. For example, some human supervision would have prevented Facebook from widely featuring fake news on the site. Some things escape AI programs, but won't escape human eyes.


However, humans aren't always able to teach AI how to avoid mistakes. When a human also tried to stop an AI from making mistakes while playing a game called Road Runner, the human overseer failed. Pong was a simple game, but Road Runner was more complicated.


Human oversight isn't going to make a perfect AI program. For another thing, we can't train AI to prepare for every possible situation it might face. However, one way to make sure that AI mistakes wouldn't have heavy consequences is to put safety guarantees. This is especially important in AI involved in healthcare and driving.



Hey! Where are you going?? Subscribe!

Get weekly science updates in your inbox!