Latent Space

Jake Elwes

United Kingdom
Student Award

Latent Space is created by Artificial Intelligence, using a deep convolutional generative adversarial neural network.

It is one of a number of pieces based on new developments in machine learning in which an A.I. algorithm has been used to generate images based on how the human brain works. Trained by inputting 14.2 million photographs, it mimics the way we learn by self-forming connections and making sense out of visual data.

Once the A.I. has built a map of neural pathways to comprehend any number of given forms within a multi dimensional space, it can then start to generate its own images as it continuously moves through the areas between and beyond what it has learnt from us. In the process the machine becomes an abstracted digital subconscious, creating a perpetual flow of dream-like images which are removed from us while also being uncannily similar to our own inner thoughts and imaginings.

Deep Convolutional Generative Adversarial Neural Network (PPGN)

Running on Amazon EC2 GPU Server (using nvidia-docker)

ImageNet - Public domain image dataset of 14.2 million photographs.

Custom shell and python scripts (using Tensorflow and FFmpeg)

Latent Space (9 minute exract)