Thursday, 2nd November 2017
How Adversarial Attacks Work. Adversarial attacks against machine learning classifiers involve constructing an input that deliberately produces the wrong classification. This article shows how these can be constructed, and includes examples generated using PyTorch which produce a sports car that gets identified as a toaster and a photo of Sylvester Stallone that gets classified as Keanu Reeves. # 8:25 pm
I’m concerned that this character will open the floodgates for an open-ended set of PILE OF POO emoji with emotions, such as CRYING PILE OF POO, PILE OF POO WITH LOOK OF TRIUMPH, PILE OF POO SCREAMING IN FEAR, etc. Is there really any need to add a range of emotions to PILE OF POO? I personally think that changing PILE OF POO to a de facto SMILING PILE OF POO was wrong, but adding F|FROWNING PILE OF POO as a counterpart is even worse. If this is accepted then there will be no neutral, expressionless PILE OF POO, so at least a PILE OF POO WITH NO FACE would be required to be encoded to restore some balance.
The idea that our 5 committees would sanction further cute graphic characters based on this should embarrass absolutely everyone who votes yes on such an excrescence. Will we have a CRYING PILE OF POO next? PILE OF POO WITH TONGUE STICKING OUT? PILE OF POO WITH QUESTION MARKS FOR EYES? PILE OF POO WITH KARAOKE MIC? Will we have to encode a neutral FACELESS PILE OF POO?