Mechanically reproduced neural networks (and other [sonic] explorations) [Multimedia exposition]

Abstract

The notion of ‘error based process’ is an area that my practice is currently exploring. Artificial Intelligence, more specifically Neural Networks are fascinating not only for the output generated from any given input but also for the workings behind the scenes: the so-called ‘hidden layers’ - the machine ‘learning’ from its mistakes. I have been making use of Google’s tensorflow and Ableton’s new probability pack to generate audio in real time (mainly in the form of MIDI) and have been gathering these motifs (currently up to 49 notes) and applying them to an open-source, parametrically designed music box printed with a 3D printer. The resulting audio is later re-recorded using contact microphones, the sound is processed in an analogue effects chain and later converted back to MIDI, in a way mimicking the AI but also collaborating with it. Later this material is used as either sample triggers, mechanical loopers or utilised to generate visuals. I have been previously working with programmable music boxes to synthesise films made up of cut out video clips where the film becomes a musical instrument, whilst at the same time the musical instrument acts as filmmaking (or remaking) device.