4 ms·
You don't need to randomly flip pixels in the input image. You can simply backpropagate to it w.r.t. the loss function and modify it based on the gradient to mi
by albertzeyer 4y ago
You don't need to randomly flip pixels in the input image. You can simply backpropagate to it w.r.t. the loss function and modify it based on the gradient to minimize the loss.
- pietro72ohboy 4y agoBoth approaches are correct. The parent is talking about Ablation based importance attribution whereas you’re using gradients to assign importance. Your approach will compute pixel importance in a single step whereas ablation based approaches generally need many passes.
- danuker 4y ago> Your approach will compute pixel importance in a single step How do you address the vanishing gradient problem?