Tripping Neural Networks live on Twitch


Jonas Degrave und Studis vom Reservoir Lab der Uni Ghent in Belgien haben eine abgespeckte Version von Googles Tripping Neural Networks gebastelt, in Form eines endlos zoomenden Bilds, halluziniert von ein paar Algorithmen, live auf Twitch: Large Scale Deep neural net falling down the rabbit hole. (via Martin)

The AI in question is an artificial neural network. It looks at images and classifies the objects it sees them into 1000 different categories. Instead of using it for classification, we are showing it an image and asking it to modify it, so that it becomes more confident in what it sees. This allows the network to hallucinate. The image is continuously zooming in, creating an interesting kaleidoscopic effect.

Note that every image generated by the network is unique: nothing is ever repeated, and all images are generated in response to the viewers’ suggestions. Nothing in the images is copied from somewhere. They are fully generated on the fly by whatever the neural network thinks makes this object this specific object.

LSD neural net – Large Scale Deep Neural Net visualizing top level features