Here, a Compositional Pattern Producing Network (CPPN) is hooked up to a trained image classifier, then we optimize the weights of the CPPN to maximize the activation of a neuron in the classifier.
In this case we maximize one of the output neurons (strawberry, lemon etc.). In addition, we impose a hue constraint on the CPPN so that the resulting image takes on a cyan tone.
CYAN STRAWBERRY
ResNet50: 100.0% strawberry
CYAN LEMON
ResNet50: 100.0% lemon
CYAN BANANA
ResNet50: 100.0% banana
CYAN ORANGE
ResNet50: 100.0% orange
The pieces here are heavily abstract, yet curiously they are somewhat representational once viewers discover the titles. There is a hint of otherworldly strawberry, lemon, banana, orange.
Each work strongly activates a label in an image classifier. In other words, an image classifier sees meaning in every image, ranging from 'strawberry' to 'orange'. These graphics represent a private language understood only by other non-human algorithmic classifiers, while often completely abstract to human observers.
Humans know that there cannot be a cyan strawberry. That a cyan strawberry is not really a strawberry, not wholly a strawberry. Yet, the algorithm not only generates a 'CYAN STRAWBERRY'. The 'CYAN STRAWBERRY' is accepted as 'very strawberry' by an image classifier, with a high confidence level.
What is a 100.0% strawberry?