Psycho AI

3

October

2018

No ratings yet.

download

Of all the technologies currently developped, few have been scutinized as closely as AI. Indeed, while the previous Industrial Revolution was about automating human tasks, the next one is likely to see machines making decisions in our places. The development of machine learning refers to a computer software that seeks to reproduce or mimic (perhaps with improvements) human thought, decision making, or brain functions (Gallaugher, 2016). This learning is conducted by feeding AIs with millions of examples labelled with the right answer. But what would happen if the wrong answer was labelled right -if the learning went wrong?

That is the question Dr Yanardag tried tackling. Instead of regular data sets, she exposed her algorithm -nicknamed Norman, after the eponymous murderer in Hitchcock’s Psycho- to the darkest corners of Reddit. Norman was then presented with a picture, for which he should find a caption. While standard AI system would label a picture ““a black and white photo of a baseball glove.”, Norman would see “man is murdered by machine gun in broad daylight.”. When the standard AI would see “a person is holding an umbrella in the air.”, Norman would see “man is shot dead in front of his screaming wife.” -for more examples, feel free to visit their website in the bibliography.
Apart from the dark humor -and certain morbidity-, this experiment underlines the importance of the initial dataset used for learning. While most people are afraid of algorithms being biased and unfair, they should focus on understanding the learning process said algorithm undertook. Norman itself was a perfectly standard system; however it learned from the worst data set possible.
Probably within the next years, certainly within the next decades, we are bound to see machines making crucial decisions in our place. To ensure optimal decision-making, our first step is to make sure machines learn from the best we can offer.
After all, machines will only be psycho if we teach them to.

Bibliography:
Gallaugher (2016), Information Systems: A Manager’s Guide to Harnessing Technology, Gallaugher.
Norman-ai.mit.edu. (2018). Norman by MIT Media Lab. [online] Available at: http://norman-ai.mit.edu/#inkblot [Accessed 30 Sep. 2018].

Please rate this

Leave a Reply

Your email address will not be published.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