Menu
AFFILIATIONS: Al ROBOLAB. Luxembourg University
AUTHORS: Daniel Gareev, Oliver Glassl
Initiation 1s a song by the ThalamusProject. The thalamus 1s a part of the brain that relays human perceptions to the human consciousness. The anatomical name is derived from the ancient Greek word 86.Aaµoc;, which is frequently translated as “gateway to consciousness”. The m us1c of the ThalamusProject directly relates to this picture: acoustic textures are used as a canvas to depict the interplay of objectivity and subjectivity. Initiation is a song about existence and unfolding reality, inspired by emanationist philosophic theories (e.g. the 4 Worlds of Qabalah). The song hereby describes the first stage of any creation, which is the pure potential, the possibility, the space, the matter and the motivation for creation. For the video of Initiation, the creators used artificial intelligence and machine learning to create a cinematic vision of the song from within the “mind’s eye” of an A.I. To create this, the authors trained A.I. on thousands of images of landscapes, artworks and photographs that depict the conceptual meanings of the song’s lyrics. As it learned, the A.I. developed neural models representing these concepts. The creators’ final edit takes the spectators through the multidimensional latent space of the A.l.’s neural models, traveling across seamlessly morphing landscapes, artworks, and photographs. The resulting film is an epic yet intimate journey across these concepts imagined by artificial intelligence.
The authors collected thousands of images representing the semantic concepts of the song’s lyrics (e.g. for the song line “I’m the first ray of light”, the images tagged with “Ray of Light” and related semantic keywords were collected). For each semantic concept, a separate set of images was assembled. The images were sourced from Bing Image Search.
This step helped to group visually similar images as coherent datasets and to identify those which represent the concepts best.
A machine learning algorithm takes a set of data known as “training data” as input. The learning algorithm finds patterns in the input data. This process is called training. The output of the training process at each step are the predictions made by the model. There are many types of machine learning models.
A generative adversarial network (GAN) is a machine learning (ML) model in which two neural networks compete with each other to become more accurate in their predictions with each training step.
The creators trained fifteen GAN models, one for each semantic concept appearing in the song. The training of each model took more than a thousand of iterations over three days. See below, how the predictions evolve.
Once the training process was complete, the creators generated interpolation loops. At the core of this step is a latent space interpolation. By interpolating between images in the latent space, the creators performed a seamless, yet true to source, transition between images generated by the model. For the final video, the authors created an interpolation loop for each of the fifteen neural models and combined them in one final cinematic film.
Input for the GAN model. Images collected for the song line “ray of light”