Closed Loop

Jake Elwes

*1993 in London (UK), lives and works in London (UK)

Collaborative project with Roland Arnoldt. Special thanks to Anh Nguyen et al. at Evolving-AI for their work on GANs.

Closed Loop

2017, 2-channel digital video, 3 hour 14 minute loop

Closed Loop is a recording of two artificial intelligence models conversing with each other – one with words the other with images – in a never-ending feedback loop. The words of one describe the images of the other, which then seeks to describe the words with a fresh image… Two neural networks getting lost in their own nuances, sparking and branching off each other as they converse in a perpetual game of AI Chinese whispers.

The piece shows two forms of neural network: a language captioning Recurrent Neural Network writing what it sees in the images generated, and a Generative Neural Network creating images responding to the words generated. The neural networks have been trained on large datasets, a dataset of 4.1 million captioned images to train a language network, and the ImageNet dataset of 14.2 million photographs to train the image generator network.

After going through the training process, during which the AI learns characteristic features of what material objects look like on a pixel basis (in images) and how they can be described using language, the neural network is then able to autonomously create images and words.