It's not maximizing "fear response", I'm not sure where everyone in this thread is getting that from. It's maximizing the response of particular visual cortex neurons, in structures where they're shown to be recognizers of specific shapes or concepts.
I.e. it's evolving images that this neuron thinks look most like a monkey, or a person. It's not the only neuron making that judgment, and it's all super nonlinear, so ofc it looks strange and distorted.
I.e. it's evolving images that this neuron thinks look most like a monkey, or a person. It's not the only neuron making that judgment, and it's all super nonlinear, so ofc it looks strange and distorted.
It's literally just doing this but with a real neuron instead of a virtual one: http://yosinski.com/deepvis