Trans-ATMs of NYC (Photography & Machine Learning)
2018-2021
Freeman Alley
In 2018, I photographed and recorded the locations of hundreds of Automated Teller Machines or better known as an ATM. The ATMs were all located outdoors on the island of Manhattan. My interest in these financial utilities came about when I photographed an ATM near Washington Square Park. The machine was blended into the veneer of the city - encased in a metal frame with openings for the buttons, slots, and a screen, retouched with a few graffiti marks. I was immediately enamored with the way in which this highly secured technology received alterations from the urban environment.
7 Carmine Street
One photo was not enough so I decided to take more photos on a mass scale. I spent days walking Manhattan, looking for every exterior facing ATM I could find. During the photographing process I recognized three demographics of an ATM: units that were either well protected, unprotected, or desecrated. Tightly controlled and monitored zones had ATMs that looked clean and safeguarded from vandalization with security personnel and cameras being present. Whereas 3rd party vendor businesses like convenience stores, laundromats, barbershops, salons, clothing stores, bars and restaurants, had machines that were functional but often plastered in spray paint, stickers, flyers, and street art. And there were defunct ATMs whose electronics were gutted, reconfiguring its hollow shell to be garbage receptacle, belonging to vacant commercial spaces or businesses that had another ATM located indoors.
Left: 960 6th Ave
Center: 132 Avenue C
Right: 8 Cortlandt Street
Center: 132 Avenue C
Right: 8 Cortlandt Street
317 6th Avenue
8 Mott Street
1 Clinton Street
23 Clinton Street
Visually, the vandalized ATMs were the most interesting category. Graffiti and the arts, which is well known in NYC, was used to make over ATMs and undermine banking institutions with alternative messages for promoting brands, services, and social issues. I initially interpreted these changes as a reactionary protest by residents feeling the economic pressures of rising living costs in New York. In actuality it is was more common that the ATM was used as a site for entrepreneurial self-promotion.
96 Stanton Street
163 Ludlow Street
32 Avenue C
These cash machines have acquired a secondary use for displaying messages. Although an ATM is built for electronic transactions and withdrawing money, applying visual information to its exterior is also a technological aspect of the machine. Techno-social theorist Bijker Weibe writes that technology is more so created by humans through a social invention than engineering. That the real character of technology arises from real world uses which come out of personal practices for utilizing technology, instead of intended manufacturing specifications or marketing propositions. Until used, ATMs are rather invisible objects. They are easily identified and found through accompanying signage that reads “ATM.”
105 Stanton Street
3433 Broadway
9 Stanton Street
On a technical level, the de-centralized operation of banking machine as a node in a financial network, enabled faster and more frequent economic activity. 24-hour access to personal bank accounts, increased profits for banks and credit lenders through transaction fees. By fixturing and renovating ATMs into the city’s architecture with stands, blocks, windows, and columns, New York could truly never sleep. During the project I learned that most users were people who lived too far away from bank branches, and used the ATM to pay their rent and bills.
2nd Ave
110 Forsyth Street
Considering the ubiquity of money, it makes sense that the machines themselves are indispensable sites for global capital. The importance of money makes the machines prime real estate small businesses. But on a macro scale it is real estate for advertising. ATM customers are forced to engage with the visual noise image, text, and painting. This mechanical dispenser that had overwritten urban infrastructure by being installed in openings like windows, doors, booths, platforms, and sidewalks, was overwritten and transformed by the activity of the city. The identity of these mass-produced machines that become a part of a whole ecosystem had changed to be unique and individualized. The ATM of NYC had become trans, as in trans-technology. They are Trans-ATMs that gradually came into the world over time.
I did not uncover this idea of trans-technology in the project until much later. When I had my collection of photos finished, I organized the images to be a book. But was the book about tourism, geography, culture, or street art? I could not decide on a topic, so the project went dormant. It wasn’t until I began training a machine learning model with the photos of ATMs that I renewed the project. In 2021, I was using Runway ML’s Lab application (now discontinued) to program pre-made models. This was during a course in Advanced Computation (DMS 605) I took with Prof. Mark Shepard, at University at Buffalo. The class was dedicated to artistic uses of AI. When I learned that the best AI models are trained on thousands of images files I recognized that I could use the collection of ATM pictures to be a dataset. I wanted train an AI model with Machine Learning to teach me something about the ATM photos that I did not know yet.
I used an object detection model called YOLOv4 (You Only Look Once), for identifying images of ATMs. I created a list of tag names and labeled parts of the digital photos with bounding boxes. Each tag category required at least 10 samples but more sampling made models more accurate. The goal for an object model is to have a high confidence score (ranging from 0 to 1.0) for knowing when trained objects appear visually. Over the course of 3 model trainings, I struggled to target what the model should focus on. Was it the ATM model names, the markings, or the urban infrastructure as whole that mattered?
Taking the long way, I made 2 overblown models with too many labels, too many results, and very low confidence scores. Consequently, the more labels I made, the work labor I had to do. As I studied more and more images, my reasoning for using or not using tags, or changing tagging was susceptible to my arbitrary judgement based on my mood and the high volume of photos containing a variety of visual information. As I trained models, the data set of photos and rules of Runways training tools also trained me.
categories":["GenMega GT3000","Village","Genmega GT5000","Tranax Mini-Bank c4000","Lower East"],"scores":[0.8740992546081543,0.6421939730644226,0.25731125473976135,0.18840231001377106,0.13273189961910248]}
"categories": ["MiniBank-1520"], "scores": [0.8502820134162903]
"categories": ["MiniBank-1520"], "scores": [0.5506561398506165]
Simplifying the process, the 3rd model used only for the manufactured ATM model names. The results were surprising in comparison to the 1st and 2nd YOLOv4 models. The number of objects detected dropped. Confidence scores increased to a much higher level than before. But in some instances, there was no result, which did not happen with the other models. No ATM was detected for both images that were and weren’t trained on the data. Some images of immaculate advertising photos ATMs had no result. And changing the hue of the photo generated a result when the original color space did not.
By training the model with photos of ATMs with different visual qualities resulted in machine that had no ID category. The sampled training data was organically inconsistent because the pictures showed ATMs in a natural condition. The ATMs had achieved undefined identities due to the plurality of their function.
"categories":["Nautilus Hyosung HALO II","Tranax Mini-Bank c4000"],"scores":[0.22653287649154663,0.11830942332744598]}
There is something amusing about a machine looking for machines, in images produced by a machine. Machines work procedurally, and procedures work best with predictable outcomes. The lack of stable identity of analyzed ATMs photographed work best if their appearance remains constant. Categorizing identity works best when things do not change. A photo can be a way to retain permanence but the same subject photographed at different points reveals a transformation.
Today, trans identities for human beings works in a similar way. The word ATM is a general descriptor like race, class, or sex, and is applied generally to all machines, no matter the size, make, year, or location. Names for what a person’s gender, sex, and orientation generate variety of names, creating an expanding the list identities. This observation is not meant to say that trans-gender people should not be acknowledged or need to have alternative labels. Rather that, like the indeterminacy of identifying a transformed ATM, it is too difficult to come up with up with a concrete label. A system for labeling images overwhelmed me due to the circumstances and rationale for using a label. Trans identities for individuals and machines show that categorization is a construct of forced processes for fitting things into organized knowledge. Like these public facing technologies, we all change and only require labeling depending on systems of governance and control.
The ATM pictures are no doubt pleasing to look at, but whose fluctuating condition is frozen as an image. Since taking these pictures, the ATMs have changed or have been removed, junking the accumulation of information contained by the machine. Just like people, there will never be another ATM quite like this. An environment formatted by technology that is invented and installed by humans, no doubt informs us about our behavior and rules. A machine that inherits our unstructured values gains personality that cannot be recognized by other machines.
About Me ︎ Curriculum Vitae
Project Index︎︎︎
2023
Videosphere: You’re Always On A Screen Somewhere...
Brown Bunny
Inertia
2022
MAX Jitter Programming
FxR Versus AhR
Async Rolling
Always On
Enframing Rey
The Vidiot
Computer Home Entertainment
2021
Machine Learning: ATM
Pursuing Earth
THE LIBERTARIAN
Wiley World Wide Weather
ReWind Compressor
Discarding Earth
Being A Passenger
2020
Can(t) See Me Now!
ColorPress Brush
White Ball Flux Brush
THE APP
2019
PATS
2018
Ice Cream Drone
2015
Vanilla ISIS
2017
Invisible Foci
2014
Digital Synthesis Illustration
2013
Rood World
Synthesis Ilustration
2011
Verse
Master (Series)
Video Being
2010
Dance Aches
Bangalor
︎ macherno@buffalo.edu
︎ videoarchaeology
︎ michaelchenoff
︎ michaelchernoff
︎ videosignified
︎ michaelchernoff
︎ Videosphere
©2010-2023 Michael Chernoff