On the limits of neural network explainability via descrambling

We study fundamental limits on explaining neural network decisions through descrambling methods, identifying when such approaches can and cannot recover meaningful structure from learned representations.