This neural network has a filthy mind
Paging Doctor Freud. An artificial-intelligence program has been made that looks at innocent pictures of beaches and deserts and sees quivering landscapes of genitalia.
Created by computer scientist Gabriel Goh, the project makes use of two neural networks. The first is a DeepDream-like system from MIT, which synthesises new images based on patterns it recognises from a bank of pictures. The second is an open -source neural network from Yahoo, which scores images on a scale of 0 to 1 based on how NSFW they are.
“Like Google’s DeepDream, this visualisation trick works by maximally activating certain neurons of the classifier,” writes Goh. “Unlike DeepDream, we optimise these activations by performing descent on a parameterisation of the manifold of natural images.”
Essentially, Goh meshed the two systems together, so that he was able to rank the generated, DeepDream-like images by how pornographic they appeared to Yahoo’s neural network. More interestingly, however, was what happened when Goh used the system to create images with a high NSFW rating, based on scenes of known imagery, such as deserts, beaches and rock concerts.
The results are pictures that look a lot like natural vistas, but have been arranged in a way that emphasises shapes appearing to be pornographic. Basically, the world becomes a seething mass of human sexuality.
Nothing new there, then.