Heteropoietic Computation: Category Mistakes and Fails as Generators of Novel Sensibilities

In the first of four domains of computational experience to be explored, the machine learning of images—by text-to-image AI, convolutional neural networks, and generative adversarial networks, for example—is critically probed. The ordinary running of these models produces extraordinary category mistakes—instances where the model confidently predicts erroneous outputs. Category mistakes become a lens for investigating the odd sensibility of machines learning. AI claims to become an engine of cultural production using style transfer, for example, to automate the capture of aesthetic genres and generically transpose and generate them. Against this, data science experiments with category mistakes are repositioned as an alternative to a homogenizing and generic aesthetics for machine learning.