In this video, I describe what the gradient with respect to input is. I also implement two specific examples of how one can use it: Fast gradient sign method (FGSM) and Integrated gradients. The first one is an adversarial attack that perturbs the input image in a way that fools a classifier network. The second is an approach that tries to interpret what features the classifier network is using to make a given prediction.

Code: [ Ссылка ]

00:00​​​ Intro
00:45​​​ FGSM method paper
01:56​​ Utility functions
07:22 FGSM example
12:58​​​ FGSM result
13:23​​​ Integrated gradients paper
14:32 Integrated gradients example
18:37 Integrated gradients result
19:13​ Cool projects + outro

Explaining and Harnessing Adversarial Examples - [ Ссылка ]
Axiomatic Attribution for Deep Networks - [ Ссылка ]

Foolbox: [ Ссылка ]
Captum: [ Ссылка ]

If you have any video suggestions or you just wanna chat feel free to join the discord server: [ Ссылка ]

Credits logo animation
Title: Conjungation · Author: Uncle Milk · Source: [ Ссылка ] · License: [ Ссылка ] · Download (9MB): [ Ссылка ]

свернуть