Thank you very much. Why they prefered the sign instead of the gradient at 7:40 mathematically if possible please? What is wrong if some dimensions go beyond epsilon please as you mentioned 7:53?
If we can easily create examples that trick AI into making mistakes, it could lead to dangerous situations. For example, someone could wear a t-shirt with a pattern that confuses a self-driving car’s AI into thinking it’s seeing a speeding truck driving the wrong way. This kind of vulnerability is a serious safety risk.
Great video Dan thanks!
Subscriber no 50.
Thank you very much. Why they prefered the sign instead of the gradient at 7:40 mathematically if possible please? What is wrong if some dimensions go beyond epsilon please as you mentioned 7:53?
How is this a safety issue?
If we can easily create examples that trick AI into making mistakes, it could lead to dangerous situations. For example, someone could wear a t-shirt with a pattern that confuses a self-driving car’s AI into thinking it’s seeing a speeding truck driving the wrong way. This kind of vulnerability is a serious safety risk.