This glorious menagerie of dick-esque space-oblongs does not contain any actual dicks. Instead, it’s the result of an image manipulation experiment. A month ago, Yahoo launched an open-sourced neural network, called “open_nsfw,” that rates images on a scale of 0 (SFW) to 1 (NSFW). So Gabriel Goh created a Google DeepDream-like visualization technique to mess with it.
Like Google’s Deep Dream, this visualization trick works by maximally activating certain neurons of the classifier. Unlike deep dream, we optimize these activations by performing descent on a parameterization of the manifold of natural images.
There are a lot more dense explanations where that came from, with formulas like this:
The quest for dick is complex, friends.
All of this potentially explains, to some of us (not me), exactly how these [not actually] dicks are conjured out of nowhere. Goh’s experiment looks at the neurons of Yahoo’s network, down to the very deconstructed visual notions of NSFW. It then combines that dataset with another neural network, MIT’s scene recognition model, called “places-cnn,” which can recognize the essence of “beach,” “concert,” and “art gallery,” for example.
It then pumps up the beach using “places-cnn” and the dick with “open_nsfw” at the same time, to discover what a beachy 0.2 on a scale of 0 (SFW) to 1 (NSFW) would look like. What do you get? A “beach” that is progressively “dick.”
Rich Oglesby of Prosthetic Knowledge tries to help me understand this great dick beach mystery.
“In the example of the beach, the beach neural layers were activated to create beach-likeness, combined with ‘open_nsfw’ neural layers,” he said. “An image was formed based on the numerical scale input. Not every output from this process will be the same necessarily. I mean, if you pressed enter 1,000 times, you probably would see some familiar features here and there.”
The quest for dick is also rigorous. But like all science, if you stick with it, over time, the truths of the universe will reveal themselves to you.