Non-robust feature means a feature which is weakly predictive (so can be useful to learn by the model) but whose value can be changed by a very small input perturbation (i.e., in the example above the Gaussians are very close to each other).
As for the yellow values, these are just the centers of the 2 Gaussians. The exact values may seem peculiar but they are chosen to simplify the derivations on the next slides.
Non-robust features in adversarial examples
Hello,
Can someone explain what does a non-robust feature mean and the meaning of the values in the horizontal axis which is highlighted with yellow?
Thanks,
Hi,
Non-robust feature means a feature which is weakly predictive (so can be useful to learn by the model) but whose value can be changed by a very small input perturbation (i.e., in the example above the Gaussians are very close to each other).
As for the yellow values, these are just the centers of the 2 Gaussians. The exact values may seem peculiar but they are chosen to simplify the derivations on the next slides.
I hope that helps.
Best,
Maksym
Add comment