I’m not quite sure this is what Yahoo had in mind when it decided to release the source for its work on a neural network designed to filter NSFW images out of its search results. Machine learning enthusiast Gabriel Goh decided to flip the switch on the code from “input” to “output” and the results… well, it gets very not-safe-for-work here on in.
You’re probably looking at the lead shot above and thinking “Hey, that’s not going to fuel my sexy, sexy nightmares!”. And you’d be right. Those images were produced by minimising the network’s NSFW score. According to Goh, this results in images based largly on the neutral “training set”, which appears to have had a thing for Monet-like landscapes.
Goh then decided to play around with the NSFW and SFW scores (called neurons in the network) by “[exciting] both … simultaneously”. That’s when it starts getting Salvador Dali:
Clearly, there’s some… stuff… going on here, but not enough to, uh, causing loin-based consternation. So what comes out when you maximise the NSFW node?
Well, to put it simply, dicks. Lots of dicks:
Goh doesn’t stop here. Who would? You should hit up the page he has showing other outputs from the neural network, including splicing it with MIT’s scene recognition network, which produced singing dicks, sand dicks and dick volcanoes.
You know, I think Yahoo’s network might have a wee bit of an obsession.