Follow
"these models are vulnerable to a kind of “typographic attack” where adding adversarial text to images can cause them to be systematically misclassified."
📄 https://distill.pub/2021/multimodal-neurons/
@liaizon
This paper explains it in a very interesting way too! With plenty of data supporting the idea of neurons which respond to concepts however they appear.
WHO WILL WIN
A neural network trained on thousands of images
VS
iPod iPod iPod iPod iPod iPod iPod iPod
@liaizon Who knew that the people wearing shirts that said "pants" were way ahead of their time?
Well I would say this is an objectively good thing at this point
@liaizon
This is hacking on a whole new level