Text-AI-Le: disentangled color directions in self-supervised deep learning models for textile image generation

As Matteo Pasquinelli puts it artificial intelligences are ‘nooscopes’, lenses that see the data through diffractions and distortions. In human efforts to understand such machines, we strive to qualify such distortions to bridge the epistemic landscape of machines’ perception to that of humans. An area of such efforts is that of latent space studies, where we project the representation of models to the vector space of a layer of the same neural network. In this setting, we explore the role of disentanglement to provide meaningful interpretative directions of this space. In the presentation, we show how textiles can be manipulated in color through disentangled directions in the latent space of models’ generative textile samples. We suggest possible avenues to interpret how the model understands colors based on such modifications.