Latent Images

Ongoing

Left: A map of E. E. Cummings’s “in Just-” inside GPT-2. Every word is a dot and you can follow its path through the layers.
Right: Halftone plate. The same points as a halftone screen. The dots are darker where hidden states are weighted by how far they jump between layers––where the model “rewrites” most.

The field of AI interpretability continually releases new tools and techniques to probe models and explain how they work. Interpretability research also generates many different types of images to represent the complex mathematical objects that are AI models: attention heatmaps, latent-space projections, feature dashboards, layer stacks, and more recently, manifolds, curvatures, and trajectories. These images are mediating objects that convey as much about the visual cultures that made them as about the phenomena they represent. Indeed, interpretability research borrows much of its imagery from other fields, such as neuroimaging, statistical graphics, cartography, and differential geometry.

These images are understood as the representations of inner operations of models, but become, in fact, the very objects that researchers study, from which they build understandings and intuitions about this new technology, which then go on to shape not only research directions but also public policies and social practices. These canonical forms remain critically unexamined, with little to no contact with the disciplines that have rich traditions of describing how representations produce knowledge: the history of science, art history, and media studies.

In Latent Images I propose a comparative media account of AI interpretability’s images, specifically the unit portrait and the scatter projection. There is an opportunity both to evaluate these images critically and to generate counter-images that draw on the same internal activations and features of models to explore the aesthetic and expressive potential of AI’s latent operations.

My project pursues two parallel objectives. First, I intend to write a critical account of interpretability’s visual culture from a comparative media perspective. Second, I will continue the work I am conducting this year, developing expressive software tools that build on existing open-source interpretability contributions, which enable researchers and non-specialists to visualize and manipulate models’ intermediate representations.

Work Home