Jump directly to the page contents


Funded by

Helmut Fischer Stiftung

Edited by

Project description

AI has advanced computer vision and image generation tremendously. But how does AI understand images? Can it connect what it “sees” to different known concepts? And when it generates new samples, is it creative? We train a system of neural networks to read images and music. Then we link it with another network that connects lines, shapes and colours into rhythm and pitch.

Organising information

Babies teach themselves how to see: they learn to distinguish colours, shapes and objects long before they know what those objects are. The autoencoder neural networks do the same: they start out knowing nothing, but teach themselves, for example, to extract, encode and decode visual information.

See the invisible, hear the inaudible

In Wolfgang Heckl’s installation “Atomare Klangwelten” (“Atomic Soundscapes”, 2018), an algorithm translates scanning tunneling microscope images, pixel by pixel, into musical notes. The microscope shows what was invisible, the algorithm plays what had no sound. Coming from a similar principle, we built a mapping between colours and notes, which turns colour pictures into music and vice versa.

Please be creative

While these pixel-note maps are intuitive, they rarely produce captivating pictures or pleasurable music. However, we can use their consistent audiovisual associations to guide the neural networks to connect pictures and music. Thus the networks learn to freely associate shapes and colours with musical ideas: they translate music samples into abstract pictures and compose short melodies inspired by images. A person doing this is undoubtedly creative. Are the machine as well?


Two-media interpolation

Images morph into one another continuously, while the soundtrack reflects the changing shapes and colours.

Other research projects