Artificial Intelligence is a very active field of research.
It is hard to believe that major breaktroughs, now fueling thriving research and applications, are so recent: AlexNet was published less than a decade ago, while Transformer is four years old, and NeRF less than two. New articles are published everyday, making it hard for researchers to follow the trends in such an intense and competitive field. New products are launched every week, making it hard for society to adapt to those new technologies, and measure their impact on our lives.
FRaGiLE is a poetic and humorous response to this frenzy. With visual dreams of futile and ephemeral objects of a mesmerizing complexity, it offers a contemplative pause in this craze of science. As a Memento Mori of Computer Vision, the short poems written in collaboration with the machine also aim at recalling how frail our beautiful architectures of knowledge are.
The creative process of FRaGiLE holds its own paradox: an "old" method – DeepDream was presented in 2015 – used on cutting-edge networks – CLIP was released in early 2021 – allows to generate eerie images from combinations of labels. This novelty too shall pass, and those pictures will one day be remnants of outdated futures.
All those images were manually optimized with state-of-the-art techniques derived from DeepDreams, based on a reimplementation of the tools described in Olah et al., 2017, augmented with Fourier positional encoding for the coordinate-based network image generators (inspired by Tancik et al., 2020), and multi-layer compositing.
Networks "deep-dreamed" this way are either the released version of CLIP (by OpenAI), or standard vision networks finetuned on custom image datasets.
The short poems presented with the images were written with the help of an apparatus similar to Write with Transformer, used with a fine-tuned GPT-2.
Cette œuvre a été réalisée dans le cadre du doctorat SACRe de l’Université PSL.