I read about analysis on local properties of neural networks. Some of them study the impact of "infinitesimal" change to an input. Like in Percy Liang's paper Understanding Black-box Predictions via Influence Functions. They have neat formalization with these properties but I don't really get the point.

My questions are:

  1. How local is local? Or maybe how small is "infinitesimal". I get the idea that this is related to some derivative and stuff. But these results are applied to some changes that are not that small from my perspective, like a change of 0.1 in pixel value ranging from 0 to 1. Why is this considered correct?

  2. From my understanding, if the derivative does not change much, it can be applied to a larger area. But is this the case in neural networks? It would be great if I can get some references on this topic.

  3. Since I don't really have much background in ml theory and serious math stuff. Other pointers that may help me understand these topics are also appreciated.

Thank you for the help in advance.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.