A starling galley of phantasmagoric images generated by a neural network technique has been released. The images were made by some computer scientists associated with Google who had been using neural networks to classify objects in images. They discovered that by using the neural networks “in reverse” they could elicit visualisations of the representations that the networks had developed over training.
These pictures are freaky because they look sort of like the things the network had been trained to classify, but without the coherence of real-world scenes. In fact, the researchers impose a local coherence on the images (so that neighbouring pixels do similar work in the image) but put no restraint on what is globally represented.
The obvious parallel is to images from dreams or other altered states – situations where ‘low level’ constraints in our vision are obviously still operating, but the high-level constraints – the kind of thing that tries to impose an abstract and unitary coherence on what we see – is loosened. In these situations we get to observe something that reflects our own processes as much as what is out there in the world.