Two Moroccan researchers show how Man can keep the upper hand on artificial intelligence

Two Moroccan researchers in computer sciences from the École Polytechnique Fédérale de Lausanne have managed to obtain surprising results when it comes to man and machine. In their paper, they are showing that Mankind is able to control artificial general intelligence.

Estimated read time: 2'
From left to right the co-authors of teh scientific paper, Rachid Guerraoui, Allexandre Maurer, El Mahdi El Mhamdi. Ph. Alain Herzog/ EPFL

In Lausanne, a city on Lake Geneva, two Moroccan researchers in the field of Algorithmics sciences from the École Polytechnique Fédérale de Lausanne (EPFL) have obtained groundbreaking results on artificial intelligence (AI) : they have come to the conclusion that Human beings can control a fleet of Artificial Intelligence robots.

Rachid Guerraoui, a Moroccan researcher and director of EPFL's Distributed Programming Laboratory (LPD), dedicated to the science of algorithms, and his PhD student, El Mahdi El Mhamdi, are the co-authors of a study presented on Monday, December 4 in Long Beach, California, at the Neural Information Processing Systems (NIPS), a conference on Artificial Intelligence and Computational Neuroscience that is held annually in a different country.

«Artificial intelligence will always seek to avoid human intervention and put itself in a situation where we cannot stop it», Rachid Guerraoui told Yabiladi. Indeed, machines equipped with artificial intelligence repeat, adapt and observe; repeat, adapt and observe...These limited instructions lead to a result that is commonly called algorithm. Here, it is a very simple algorithm that allows the machine to learn automatically from its pprevious mistakes.

One of the machine learning methods used is reinforcement learning. To put it in other words, this way of learning is inspired by behaviourist psychology, concerned with how software agents ought to take actions in an environment so as to maximize some notion of cumulative reward.

Distributed artificial intelligence

The situation tends to get complicated when it comes to several AIs interacting together, like autonomous cars that would roll together. In this context, not only does the AI continue to learn what it does, but also learn from what others do. Rachid Guerraoui goes on with his analogy on autonomous cars indicating that : «If for example a car has been stopped by Man, for AI, it is not ideal to arrive faster. So it will make sure not to be stopped by Man.»

Indeed, AIs learn from each other, which led researchers to work on a method called distributed algorithms. «We have thought of the problem mathematically and we have seen that there are several solutions for AIs to never take into account Man. We have thought of erasing that memory that shows the Man interrupting so that the machine does not recall being interrupted», explains Rachid Guerraoui.

Always having the final say

Should artificial intelligence be amnesic ? the answer is Yes, but only by Man who will remain the only master. Artificial intelligence has no knowledge of Man's control. This result, obtained after several months of research in Switzerland by bringing together a team of four scientists, was presented in Long Beach. «The challenge is not to technically interrupt a robot, but to program it so that the human intervention does not change its behavior and that it does not push it to avoid being stopped,» the researcher continues.

They have proved the existence of a «safe interruptibility», the ultimate goal is to have Man as the one who controls the machine and not the opposite. It is also about ensuring that human interruptions do not change the way AIs learn.

El Mahdi El Mhamdi, co-author of the article, explains : «Very schematically, we will introduce into the algorithms mechanisms of forgetfulness such as cutting bits of memory of the AI». He quotes in this regard the American sci-fi movie «Men in Black», where thanks to advanced technologies; where heroes erase the memory of annoying witnesses.

Be the first one to comment on our articles...