"these models are vulnerable to a kind of “typographic attack” where adding adversarial text to images can cause them to be systematically misclassified."
📄 distill.pub/2021/multimodal-ne


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.

Sign in to participate in the conversation
La Quadrature du Net - Mastodon - Media Fédéré

Mamot.fr est un serveur Mastodon francophone, géré par La Quadrature du Net.