One of the most interesting results from “Explaining and Harnessing Adversarial Examples” is the idea that adversarial examples for a machine learning model do not arise because of the supposed complexity or nonlinearity of the model, but rather because of high dimensionality of the input space. I want to take a stab at explaining “Explaining”’s result with an economics analogy. Take it with a grain of salt, since I have little to no formal training in either machine learning or economics. Let’s go!
The one-sentence version of this post: There are economies of scale with dimensions when constructing adversarial examples.