Income
Health
Education
Generated images from a model trained on photographs from London. The sliders change the features of the images according to what is typical for deprived or privileged areas.
Explanation
Generative models such as DALL-E or Generative Adverserial Networks (GANs) are primarily known as accessible, low-cost ways to generate images. Could they also be used to study social phenomena, such as the visual aspects of social processes in cities?
This inequality does not exist is a demo of a generative models, like others in the refrain of This X does not exist. It displays the outputs of a GAN model that was trained on tens of thousands of photographs of housing in London, sourced from Google Streetview.
How it works
We connected the photographs from Streetview with data from the Indices of Multiple Deprivation, describing levels of income, education and environmental health across the city. The resulting GAN model generates novel, synthetic images of housing that can be manipulated according to these contextual variables.
By pulling on the sliders next to images, you can compare the visual differences between privileged or deprived areas of the capital. It is a way to visualise an abstract social process, inequality in the city, that would otherwise be difficult to see.
A technical paper and a poster explain how the model works in more detail.
This inequality doesn’t exist is an experiment in trying out how models such as GANs can be used to produce knowledge. Latent space walks suggest a speculative configuration for knowledge production that relies on synthetic data and retains a role of interpretation by human participants. This process demonstrates the variety of uses that mathematical concepts and technical systems such as the latent spaces of generative models can be put to.
Who made this
The project is authored by Aleksi Knuutila. Aleksi Knuutila is an anthropologist and a computational social scientists working at the University of Helsinki and the University of Oxford. His work focuses on combining machine learning with qualitative research workflows, to enable new forms of investigations.