This is the underlying intuition behind adversarial attacks and defense strategies in neural networks. When we understand what patterns maximize certain activations, we can also understand how to manipulate inputs to fool the network or how to make our networks more robust against such manipulations.