Sagan

Self-portraits by Machine Learning

Hopeful Monsters, 2019

Self-portrait, 2019

Sheltering-in-place: Self-portrait with N95 mask, 2020

Sheltering-in-place: Self-portrait as Superman, 2020


SAGAN is a set of self-portraits generated by Machine Learning. For the self-portrait, 750 photographs are taken of the artist. In each photograph, the artist has a different pose, a different expression, and sometimes, different clothes. The 750 photographs are then used to train a Machine Learning neural network. The neural net is tasked with producing an image that looks exactly as if it came out of the set of 750 photographs. The resulting image is not an exact copy of any one of the 750. The goal is to produce an image that not only looks as if it belongs to the set, the image should be indistinguishable from any other photograph as being from the set.

The Machine Learning algorithm used to produce these images, SAGAN, is described in the paper Self-Attention Generative Adversarial Networks by Han Zhang, Ian Goodfellow, Dimitris Metaxas, and Augustus Odena, published in 2018. The code for this implementation of the algorithm was written by Junho Kim.

Hopeful Monsters is the output of sixteen different training cycles of the neural net.

Self-portrait is the training output which the artist deemed "closest" to being one of the 750 photos.

Sheltering-in-place: Self-portrait with N95 mask is the output of a neural net trained on 100 photographs of the artist wearing a N95 mask. This is the output of 4 training cycles out of the 55,000 cycles run.

Sheltering-in-place: Self-portrait as Superman is the output of a neural net trained on 800 photographs of the artist dressed as Superman. The image is a reaction to our society's inability to cope with a medieval plague despite having such advanced technologies as Machine Learning. The image is an homage to the painter, Mel Ramos.