Exhibition loss=“binary_crossentropy” 
Artists Juliette Penelope, Romain Biros
London Enclave Lab 17-22 December 2019 

loss=“binary_crossentropy” is an installation by Juliette Penelope and Romain Biros developed in conversation with Rita Aktay. It shares the machine learning process of a GAN (Generative Adversarial Network), configured* and trained by Romain on Juliette’s archive of medical imaging.

An extract from Juliette’s broader research into the semiotics, operational contexts and reconfigurations of technical images, the dataset for loss=“binary_crossentropy” looks at the clinical visualisation of the body under gynaecological procedures such as CT scans, hysteroscopy, transvaginal ultrasound, MRI and 4d ultrasound. It is composed of 415 images taken largely from dedicated Instagram accounts as well as medical reports.

Our GAN has trained on the dataset for a month, trying to emulate how these technical apparatuses portray the uterus. As a game played between a Generator and a Discriminator, where the former produces images and the latter evaluates their truth according to the dataset, the GAN is forever learning. The better the Generator fools the Discriminator, the stronger the images echo their object of knowledge.

A curious child and a sly trickster, the GAN incessantly refuses to agree, whilst offering always more, other and something else. In contrast to diagnostics, anomalies and cures, the undisciplined machine generates successful failures, never negating but always diverging.

Bringing the latent space of machine learning into the gallery space and thus conflating ways of image processing, loss=“binary_crossentropy” explores sense-making — human and machine, medical and intimate, scientific and indulgent. For any attempt at understanding cannot but fail, the latencies in the GAN’s failure feel more real than our supposedly unfailing normative apparatus.

Through the instagram account “xrct_001”, the resulting images are fed back into their initial space of extraction, re-indexed using specific labels such as #Advanced_Radiology #uterus #ML #eyesofmedicine #operationalimage #CTTechnologist and #medicalimaging.

* Based on Karras, Tero et al. “A Style-Based Generator Architecture for Generative Adversarial Networks.” CVPR (2018).