[monochr]

I have nothing intelligent to say without reading the full paper...

...But, how different is this from the various optical illusions humans fall for? I mean we can't exactly tell the difference between a rabbit and duck ourselves[1] so isn't it just a universal property of all neural-network like systems that there will be huge areas of mis-classifications for which there hasn't been specific selection?

[1] http://mathworld.wolfram.com/Rabbit-DuckIllusion.html

[praptak]

I believe that the point of the article is that the triggers for optical illusions are totally different in humans and the ANNs. I don't know how valid is this statement - humans sometimes do recognize "shapes" in white noise too.