This just seems true by definition. If you have a learning algorithm that has generalized well, then adding noise to its input shouldn’t result in a large degra

This just seems true by definition. If you have a learning algorithm that has generalized well, then adding noise to its input shouldn’t result in a large degradation in accuracy.