Skip to content

Comment on Neural Network Follies (2003)parent

Comments

Both 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.

Your approach will compute pixel importance in a single step

How do you address the vanishing gradient problem?

AboutSource Built by g1lg1l

Hackerly is an independent reader for Hacker News, built on the public HN API. Not affiliated with Y Combinator.