LOSS SURFACE 01

LOSS SURFACE by James Bloom visualizes the statistical output of a deep neural network as it analyzes a dataset of images for visual recognition. The artwork makes visible the process of optimization that the model employs as it undertakes the task of image classification, a process that is normally invisible.

The loss surface is a 3D graph of a ResNet 56 model’s performance. Where the model is unsuccessful, the loss function is high, producing a mountain. Where it is successful, a valley appears, representing low loss. This landscape captures a moment in the process of visual interpretation, frozen into a monolithic object with a physical presence.

A problem with the encroachment of AI into the classification of the world is that it generates a landscape in which people are alienated. As John Tagg put it, these systems "instantiate a machine regime indifferent to the human ego, eye, and body." LOSS SURFACE investigates the act of seeing and the impulse to classify and interpret.