Genesis

AvantGarden: Sphere4 Genesis by Tanja Vujinovic
Virtual reality art installation published on VRChat
3D graphics, world building, digital sculptures: Tanja Vujinovic
Sphere4 ₲Ɇ₦Ɇ₴ł₴ AV set used for training of deep neural network by Tanja Vujinović (visual material from Universal Objects and AvantGarden series) and Sasa Radic (sound);
Sound: Tanja Vujinovic/LUZ1E
Training of deep neural network: Dr Vid Podpečan, Department of Knowledge Technologies, Jožef Stefan Institute
Coproduction: SciArtLab, Jožef Stefan Institute
Production: Ultramono, 2021

Consulting
Dr Jelena Guga, researcher
Thimster, Graphics design master
Dr Vid Podpečan, Department of Knowledge Technologies, Jožef Stefan Institute
Ivan Stanić, curator and artist
Derek Snyder, researcher and editor
Friends from VRChat

Sphere4 Genesis is a virtual garden that enables us to immerse ourselves into an ongoing soundscape generated with the help of custom-made artificial intelligence software. A deep neural network was trained to generate sound through samples of minimal techno music. This world contains machinic and biomimetic-inspired objects that developed sentience through machine learning. The world is full of orchids, snakes and generated snake-like objects – AvantGarden machines, while the Dyson sphere-inspired objects float, encircling stars and capturing their energy to be used in the world. As Félix Guattari puts it, “its chaosmic Universe can be constellated with […], vegetal, animal, cosmic or machinic...becomings.” As new planes of the AvantGarden unfold, we get to know their elements, which develop lives of their own. The world of Sphere4 Genesis is published on VRChat, an amalgam of a computer game and a social platform that offers its users possibilities of organizing events and publishing content in the form of virtual worlds and avatars. VRChat could also be described as a “Non-game game”, a class of software offering the player unbound possibilities of freeform play, identity and a great degree of self-expression, exploration and interaction, without the limits of conventional or imposed goals and objectives. All Sphere4 worlds published on the VRChat platform contain “seeds” of the AvantGarden ecosystem, with the addition of newly developed ones made especially for Sphere4.

Technical description (written by Dr Vid Podpečan)
Our sound generation uses end-to-end neural audio generation model called SampleRNN. This is a recurrent neural network which generates one audio sample at a time and uses a hierarchy of modules, each operating at a different temporal resolution. In particular, we use the PRiSM's implementation of SampleRNN in TensorFlow 2 (https://github.com/rncm-prism/prism-samplernn). The input audio data which is a collection of minimal techno soundtracks is downsampled to 16kHz and sliced into 8 second chunks which are then used to train the network. The hyperparameters of the model were optimized and one of the best combinations was used to train the final model. During the training for 250 epochs intermediate models were stored at every 10th epoch. The model starts to generate useful audio data after approximately 60-100 epochs so we used all available useful intermediate models to generate audio samples while also varying the sampling temperature parameter.

References.