technews-editor@acm.org
Date: Mon, 10 Dec 2018 11:36:58 -0500
Linda Geddes, BBC News, 5 Dec 2018 via ACM TechNews, 10 Dec 2018
Computers can be tricked into misidentifying objects and sounds, raising
issues about the real-world use of artificial intelligence (AI); experts
call such glitches `adversarial examples' or `weird events'. Said the
Massachusetts Institute of Technology (MIT)'s Anish Athalye, ``We can think
of them as inputs that we expect the network to process in one way, but the
machine does something unexpected upon seeing that input.'' In one
experiment, Athalye's team slightly modified the texture and coloring of
certain physical objects to fool machine learning AI into thinking they were
something else. MIT's Aleksander Madry said the problem may be rooted partly
in the tendency to engineer machine learning frameworks to optimize their
performance on average. Neural networks might be fortified against outliers
by feeding them more challenging examples of whatever scientists are trying
to teach them.
https://orange.hosting.lsoft.com/trk/click%3Fref%3Dznwrbbrs9_6-1d7a4x219197x069560%26