After training a neural network to decide whether an image is of a dog, cat, fox, or boat, we can do various manipulations to try and understand how it works. One method is to take a single class and optimize an image that would maximize the network’s conviction of that single classification. For example, we could optimize to generate an image that isn’t just classified as "fox" at 80% confidence, it’s 99.9% fox. For some reason, the result is not foxy at all, it’s a big pile of random noise pixels. We’ll discuss why that is, how to fix it, and other techniques to improve our understanding of what the neural network is thinking when it’s classifying. Maybe we’ll find the foxiest fox after all.