Modules
All technical revolutions lead to social transformation. Nowadays we are witnessing yet another of these revolutions: that of Machine Learning and its trained models that are files “taught” to recognize patterns that serve to explain our reality. In this way, what will be regarded as human history is currently mediated by digital algorithms. Yet these use largely biased data-sets, plaguing this mediation with colonial remnants. We can ask, of course, if we should we be supportive of this merciless quantification of reality or not, but, at the very least, we need to take a critical stance. Hence the necessity of other data, thus fighting against the monoculture that produces models that can be "opinions embedded in mathematics” because these algorithms are now turning into tools for “epistemicide”: the destruction of existing knowledge, in this case, through the filtering out of non-western voices from the informational mainstream. As smaller tributaries of a wider river, non-western collective memories tend to flow on arid sides, slowly evaporating while the main body of liquid information gets mightier. In the case of Artificial Intelligence, Machine Learning models learn from patterns from data that one feeds the system which reinforces postcolonial domination, so this research project is meant as a way to fight back.
The Singing Rivers project uses the analogy of the "talking river" (as the Rimac river in Peru is known, where such Quechua word means "speech") so to multiply these voices by training popular Machine Learning models with datasets that prioritize other peoples and cultures, in this case, linked to the Amazonia and its rivers that have been the main communication channels in the region for centuries. One can access its different research branches below:
For the "Hidro" module, two different models have been used: DAIN, or "Depth-Aware Video Frame Interpolation" (Bao, Wenbo and Lai, Wei-Sheng, and Zhang, Xiaoyun and Gao, Zhiyong and Yang, Ming-Hsuan) and RIFE, "Real-Time Intermediate Flow Estimation for Video Frame Interpolation" (Huang, Zhewei and Zhang, Tianyuan and Heng, Wen and Shi, Boxin and Zhou, Shuchang) both serving to recreate missing frames in a set of consecutive images,
The "Homo" module pertains Amazonian inhabitants, and how we rarely find their characteristics appearing in popular GAN models. A dataset with various images from faces of tribes and communities that currently live in the Peruvian section of the Amazonian Rainforest has been used to train a pre-trained StyleGAN 3 (Tero Karras and Miika Aittala and Samuli Laine and Erik H\"ark\"onen and Janne Hellsten and Jaakko Lehtinen and Timo Aila) Neural Network model.
In the case of the "Textile" module, a pre-trained StyleGAN 2 (Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, Timo Aila) model has been fine-tuned on a selection of images from Amazonian textile patterns.
