Mirror Exercise (2020) is an AI-generated duet with myself. Real motion-capture data of myself dancing is accompanied by those same movements imitated and distorted through the lens of a deep VAE model I’ve trained on my own movements. This model was created alongside my collaborators Chase Shimmin, Douglas Duhaime, and Ilya Vidrin (arXiv:1907.05297 [cs.LG]). The VAE learns how to efficiently represent fixed-length sequences of movement of about 6 seconds at a time. By adding varying amounts of random noise to the encoded point representing each real sequence in the latent space, I ask the model to deviate slightly or significantly from my original phrases in its generated outputs. The title plays upon the role of imitation and repurposing in creative inspiration, as well as the common “mirror exercise” game to help pairs of dancers generate new material together, in which the goal is to move in tandem while obscuring which dancer is leading or following. In this piece, the generated dancer expands upon my real choreographic impulses, creating gestural counterpoints out of subtle alternate realities.