Hacker News Reader

If I'm understanding you, you're saying that this is more of a way to evaluate whether a network has generalized rather than a method to help it generalize. Is that right?

View on HN

This paper basically proves that neurons which people understand better are not any more "important" to the network than the ones we do not understand as well. It is absolutely true that a generalised network will have much lesser drop off in accuracy as compared to a network which basically just stored the input data.

Is this a metric that is being used any where?

Would be interesting to have 'resilency' metric in addition to the normal 'accuracy' and 'loss' metrics.

Reply on news.ycombinator.com