On the hunt for a small exhibition space in Brooklyn, I surfaced on Morgan Avenue to find myself face-to-face with a pair of small water towers emblazoned with a Super Mario mystery box.1 Despite my industrial surroundings, Google Maps assured me that my destination was nearby. I pocketed my phone and overshot my destination, a dimly lit hallway whose naked interior finishes sat beyond a pair of seemingly locked double doors. Upon closer investigation, I discovered “Energy Flows Marguerite Humeau” written in small text on one side of the hallway.
Eight creatures are staged across two dark rooms, their botanical appendages frozen in various empathetic acts. Some pieces resemble a dialogue between two or three rippling tendrils, while others exhibit a choreographed assemblage of bronchioles, buds, or bell peppers. Riffing on John Koenig’s Dictionary of Obscure Sorrows, Humeau uses sculpture rather than words to embody common “states of existence”2 —sensibilities that remain just vague enough to evade a proper label. A piece titled Yuyi—a term directly lifted from Koenig’s lexicon—connects the feeling of “wishing you could see things with a fresh eye” to the plant species Fumaria, whose morphology resembles blood vessels and “supports blood activation.” Similarly idiosyncratic pairings are made for each piece, as the artist attempts to encapsulate equally blurry phenomena.
The intersubjectivities that make up these marginal aesthetic categories reflect the world they are produced within. The way we perceive things is structured by our sociality, and so is the way that we articulate these perceptions. Cultural theorist Sianne Ngai describes this as the double-sided nature of aesthetic categories, an objective perception of style coupled with a subjective evaluation of judgment.3 She also describes marginal aesthetic categories as being “dispersed,” due to a “constantly shifting spatial or temporal reference as well as a degree of institutional codification.”4 The categories included in Energy Flows can be understood in a similar fashion; we remain without a formalized method of communication for these phenomena, and they do not have any outward spatial or temporal register. Free from both style and judgment, they remain aesthetic categories as far as they generate an affective human subject.
Considering the joint role of both perception and communication in the production of aesthetic categories, these fuzzy sensibilities provide an advantageous tool for indexing reality. The codification of cultural modalities eventually makes it difficult to distinguish their ability to index an authentic reality. As a field of cultural production sharpens, its reflection of a societal origin point grows foggy.5 The intrinsically blurry nature of our marginal aesthetic categories allows certain phenomena to remain free from what sociologist Pierre Bourdieu calls the “field of restrictive production.”6 Humeau’s pieces offer totems that represent ontological associations, however, they still rely on an accompanying written description. How do we then tap into these latent categories to gain a clearer picture of society? What do our cultural artifacts look like if we can find a way to communicate this minutia?
Humeau’s work offers us the trouble at hand: we articulate societal categories through the strictures of natural language, yet we perceive things through the synapses of our perception. Media theorist Lev Manovich describes this as a “complementary system” between top-down and bottom-up analysis, respectively. While cultural phenomena are a product of a continuous gradient of parameters, we have a tendency to segment them in our production of an archive.7
Cultural analytics (an area of research that Manovich has written extensively on) offers machine learning (ML) as a cultural modality that might begin to rewrite this system. Where Ngai describes the construction of categories as an equal divide between perception and articulation, Manovich poses a paradigm that suggests perception to be intrinsically linked to articulation; “numerical measurement of cultural artifacts … [produces] a language closer to how the senses represent analog information.”8
This collaborative method of seeing the world can be understood in the processes that undergird supervised and unsupervised learning in artificial neural network architectures that power machine learning. In both cases raw data (a training set) is fed into a machine in order to identify features between similar categories of information. In the case of supervised learning, data is categorized before features are detected, whereas unsupervised learning takes unorganized information and creates categories based on the features that are learned. It takes human perception to label the data at the beginning of the former process, and it takes human perception to validate categories at the end of the latter. An unsupervised ML model will start to cluster vague affinities between the raw data points it receives, and in the case of generative adversarial networks (GAN) it will generate new data to match.
While it is fairly easy to foreground the political and social utility of societal categories, the continuous gradient of a dataset has a way of presenting itself as objective. In the same way that a Rotten Tomatoes score—comprised of a collection of binary responses to content—presents itself as a fine-tuned percentage, the metadata of artificial neural network training sets are subject to discrete and often biased labels. This issue is exacerbated as training sets increase; as the need to furnish data with metadata requires a wider net to be cast, the spectrum of “ideologies, semiologies, and politics from which they are constituted” also widens.9 On top of assumptions being made by outsourced image labelers,10 there are also a host of assumptions intrinsic to the process itself.11
What can we learn when we compare the New York zoning boundaries to a database of streetscapes or resident descriptions? When I type in Williamsburg, Maspeth, Bushwick, and Greenpoint, the boundary that Google Maps draws for each neighborhood fails to include the location of the exhibition space I eventually found my way into.12 This might be a mapping error, a reflection of a value system, or both. If we were to draw an ML neighborhood boundary, what would that look like? Machine learning offers not only a higher fidelity depiction of the context that surrounds us, but a new contextual plane altogether—a latent value system that otherwise operates in silence, similar to Humeau’s sculptural hallucinations. Machine learning models offer us a reflexive tool, as well a generative tool; a method for thinning the fog, and expanding it further. A collaborative analysis of the context within which we are wielding our cultural practices offers the potential to simultaneously assess the legitimacy of preconceived societal categories and discover new ones altogether.
Christopher Pin is an M.Arch I candidate (’23) at the Yale School of Architecture.
- Truly, the perfect symbol for M1 zoning in New York. Truly, Core 4 studio shadowing my every thought. ↩︎
- The description of the exhibition, printed on three pages of stapled copy-paper and stacked by those unassuming glass doors, read “Conceptual maquettes for unnamed states of existence.” Author unknown. ↩︎
- Subjective aesthetic judgements are “codified ways of sharing our pleasure and displeasure with others,” they “produce a kind of illusion of apparitional quality at the level of rhetoric, analogous to that of style.” Sianne Ngai, “Introduction,” in Our Aesthetic Categories: Zany, Cute, Interesting (Cambridge: Harvard University Press, 2012), 38–41. ↩︎
- Ibid., 30. ↩︎
- “The more established the field becomes, the less can the production of the work of art, of its value but also of its meaning, be reduced to the sole labour of an artist—who, paradoxically, increasingly becomes the focus of attention.” See Pierre Bourdieu, “Historical Categories of Artistic Perception,” in The Rules of Art (Stanford: Stanford University Press, 1992), 295. ↩︎
- “ …it is the field which constructs and consecrates.” Ibid., 244; emphasis added. ↩︎
- Lev Manovich, Cultural Analytics (Cambridge: The MIT Press, 2020), 164. ↩︎
- “This language is closer to how the senses represent analog information. The senses translate their inputs into values on quantitative scales, and this is what allows us to differentiate among many more sounds, colors, movements, shapes, and textures than natural languages can describe.” Ibid., 154. ↩︎
- Kate Crawford, Trevor Paglen, “Excavating AI: The Politics of Images in Machine Learning Training Sets”, 2019, https://excavating.ai/. ↩︎
- A well-known example is Amazon Mechanical Turk, a service that allows businesses to outsource virtual labor, like the labeling of unstructured datasets. ↩︎
- Assumptions made regarding an equivalence between concrete and abstract nouns, the unchanging and universal nature of concepts that lead to applied labels, and the assumption that concepts can be expressed through the given medium, to name a few. For a more detailed description of the “foundation of unsubstantiated and unstable assumptions,” see Kate Crawford, Trevor Paglen, “Excavating AI.” ↩︎
- C L E A R I N G is in Williamsburg, according to the “Brooklyn Neighborhoods” Wikipedia page; “East Williamsburg,” according to Google Maps; and “Bushwick,” according to residents of an older generation. ↩︎