117 - Interpreting NLP Model Predictions, with Sameer Singh
NLP Highlights
00:00
The Importance of Manipulating Gradients in Computer Vision
Even the gradient on a very local level can be controlled and manipulated in a way that allows someone to manipulate what the gradient might look like. So this is not unique to gradient we've done some work in collaboration with HEMA. Lakeratu from Howard that we showed that even things like line and sharply values can be manipulated by add the city. That sort of brings into question how manipulatable are these kind of adversarial classifiers it's again a good research question.
Transcript
Play full episode
