Quick, look at the byline below the title of this post. You recognize a little me, right? That simple task took billions of neurons in your brain a mere instant to figure out. Though our computer recognition software is good, and getting better, no tech can yet compete with the balls of jelly and few kilos of meat sitting in your skull.
Case-in-point, here’s what happens when Google’s image recognition programs are primed to find signals in the noise:
The acid trip you’re looking at is the result of a feedback loop occurring in an artificial neural network, which is a series of algorithm “layers” acting like neurons do in the brain. These networks can learn — each neuron connection can adjust its own “weighting”and prioritization — and with practice become much better. Google does this with their image-recognizing neural networks by feeding them millions of photos and adjusting the layers until the output is correct. The ultimate goal is to have a program that could see any kind of picture with a dog in it, for example, no matter the lighting, orientation, color, etc., and tell us there’s a dog in the photo.
For these trippy images however, Google engineers decided to look into what each neural network layer was doing as it looked for certain image aspects, starting with basic components like edges and lines, and ending with “dog”. For example, this is what it looks like when primary layers look for edges, and then engineers tell it to find more edges, to enhance and form a feedback loop:
Image: Zachi Evenor; Günther Noack
Or sometimes Google engineers would ask a layer to key in on animal-like shapes, then enter a feedback loop of more and more animals:
Google calls this technique of diving deep into the abstractions of images at certain layers of image recognition “Inceptionism”.
Even if the engineers gave the networks just random static, but keyed them into building-like shapes, the result was just as abstract:
It’s not exactly what we’re looking for in an image recognition program, but these bizarre and often beautiful images can help programers better understand how these neural networks learn, and how to improve that process. And make horse-riding acid trip dog knights.
For more detail on Google’s Inceptionisms, head here.