March 28, 2022
[RECAP] A Week With Frederik de Bleser
Although I have had the pleasure of writing for Milieux for more than a year, I was struck by the fact that Frederik De Bleser’s workshop was the first Milieux event that I have attended “in person.” I can speculate that I was not alone in this sudden feeling of embodiment, as this series of workshops also marked the return of in-person events for both Machine Agencies and LePARC. Thus, prior to the first word, in the bodily preamble to De Bleser’s talk, a series of terms presented themselves for reflection: locality and distribution, presence and distance, body and image. A series of threads woven between this definitive pair of the digital and the physical.
De Bleser’s work often plays in the translation space between such pairs. At the core of his Day 1 workshop was a performance piece entitled CyberSensuality (created during LAbO Summer School 2021 by Nikola Scheibe, Alexandra Fraser, Madina Mahomedova, Mazarine Haarscheer and David Bello Arcos, with the help from De Bleser as part of The Algorithmic Gaze research project), which uses computer vision and machine learning techniques to analyze the dancing body and reproduce it as a computationally-generated spectral figure. This operation begins with the recording of a visible and luminous body, dancing, on its own in front of a digital video camera. After this initial moment of capture, each frame of the video is subjected to a “pose-detection” algorithm, which extracts the position of the body in space.
In this new image, the body as a visible object withdraws into a uniform blackness, and the primary abstraction of its location in space appears in the form of lines and dots. It is in the translation between these two images, the body and its geometric shadow, that the CyberSensuality algorithm learns a logic of vision and movement. Through a process of trial and error, it begins to understand how a specific clustering of dots in an infinitely dark space calls for something that we would recognize as an arm or a leg. From here, it gains the ability to drape body-like images over the skeletal frame, enveloping the formal schema with the remnants of bodily-movements. Thus, when played in a sequence at the appropriate frame-rate, the algorithm begins to approach something like dancing.
After a morning spent tracing this process, the afternoon was dedicated to recreating it. I had not intended to do any dancing during this workshop. I’m not sure whether it was the beautiful performances by various LeParc members or the generalized excitement and encouragement of being in a space with others, but either way something had shifted; I was inspired and ready to let loose. After requesting some (enthusiastic) techno, I stepped in front of the camera and stage lights and readied myself to engage in some free form bodily movement. As the music began, a series of thoughts flooded my mind. First, I felt a certain kinship with the algorithm. With no formal dance training, and under the observation of a group of people dedicated to the performing arts, I found myself calling to mind the impressions of poses I had just seen, rather desperately attempting to mimic the more graceful and measured movements of the prior dancers. And just as my singularly awkward and creative gestures were the result of this imperfect repetition, perhaps the strange morphology of CyberSensuality’s movements could be captured in the space of its loss function, which records the algorithm failure to map the input image to the output image. But just as this thought emerged, I felt my body shake with the kind of nervousness that results from the experience of raw bodily exposure. My breath and muscles tightened, and then loosened as I began to settle into the music and the fate of my movements. Since then, I’ve often wondered whether I ever stopped referencing the impressions of other movements, thumbing through an archive of gestural space recorded and remembered, reflectively articulating that schema of possibility that maps one movement to the next.
In posing the problem in terms of time and sequence, I have already drifted from the reality of the algorithmic systems used for CyberSensuality: it has no need of time in that sense. It does not capture any information about how dots and lines transform over time. It functions synchronically, mapping between a single image of dots in space and the frozen image of the body, already decomposed into frames by the technologies that capture its movement. When asked about this lack of time-representation, De Bleser noted a reality that has made itself felt in the revolutions of Big Data and Deep Learning: more data always beats more sophisticated algorithms in terms of effectiveness. In other words, if we wanted a more “accurate” CyberSensuality, our time would be better spent amassing images of bodies rather than tweaking the formal complexity of the system.
But perhaps it is the limited and local scope of an algorithmic process like CyberSensuality that identifies a key vector within De Bleser’s calls to “democratize creative AI.” The second day of the workshop, hosted by Machine Agencies, focused on the conditions and methods that define the contemporary field of machine learning: image recognition, natural language processing, sentiment analysis. In contrast to the industrial scale of data collection, which aims at predictive accuracy, tools like De Bleser’s Figment and GanDelve implicitly challenge such totalizing epistemic ambitions. By lowering the barriers to interaction, programs like Figment make the algorithmic processes that articulate our world more accessible and transparent, de-mystifying AI while intensifying the feeling of its magic. But they also subtly reroute the guiding ideals of machine learning towards a different magnitude and rhythm of practice. In a world dominated by homogenizing models with billions of parameters, owned and used by corporations for profit, this shift amounts to a re-framing of algorithmic practice as a medium for local creative expression, operating at the scale of the artist. Just as the body dancing is freed from the strife of labor, so machine learning calls for other uses. Joud Toamah’s piece 100 Hundred Days, produced using a data-set of only 100 images which were repetitively drawn from the memory of a specific place, exemplifies this use of AI to express locality rather than totality, calling back to the particularity of body and site.
At the end of Day 2, as I ran the images of myself dancing through Figment, I was struck by the intimacy of the data-set. Such an intimacy, between the self and its algorithmic capture, must be recognized as truly commonplace given the radical extension of machinic recognition that pervades our daily lives. However, it is rare that the apparatus of such processes becomes transparent and controllable, becoming manifest in their attempts to know us. Here, algorithms become the medium of local reflexivity and memory, hallucinatory microscopes into the patterns hidden in our hand-made input, fractal projections of limited rule-sets borne out of the intertwining of computer and body, logic and image.