(Deep) Representation: a Layered and Interconnected Mediation

Representation, in its most rudimentary sense of ‘standing for’ something else, operates as a fundamental yet profoundly complex mediation within contemporary AI models. These systems do not simply process information but re-elaborate it through a sequence of transformations, producing what we call deep representations: encodings that are simultaneously layered, in- terconnected, and polysemic. Drawing on Stable Diffusion as our case study, this work traces the representational journey from world to model, exam- ining how each layer of mediation reshapes what ‘standing for’ means. We begin with the mediations that precede the model: data as capta rather than given, digital image as cultural object and computational substrate, dataset as epistemic border, and pre-processing as structural constraint. We then turn to transformations within the model itself: the training, loss functions, and material economies that shape learned representations, differing fun- damentally from standardized compression. We conclude by questioning whether the inherited vocabulary of representation remains adequate to describe what these systems actually do.