Dear museums and colonial nations of the Global North, Your data is haunting you! Demons are waiting for you in latent space, ready to attack at any time and from any vector!
Digitization has engendered a notion that certain museums may become Datenkraken, hoarding datasets relating to the holdings (whose physical monopoly they already possess). The practice leads to moments like this one, where I am compelled to mention the British Museum [sic] when I digitally publish or re-mix an artefact from Iraq (such as a 3D dataset of the beautiful Lamassu, which you see here).
Yo, maybe you can already tell where this is heading? Let me put it to you straight: We live in a post-digital world as much as a post-colonial one. In the end, the important debate is not about what kinds of licenses should be granted, but about who owns cultural data and who controls it (and, perhaps, to whom it is attributed).
When I write that imperial museums are haunted by data, I am not referring to abstract ghosts, but suggesting that you abolish yourselves! Frankly, as long as your collections are not updated, you might as well be ghosted by your conscious audience.
There are so many self-appointed “universal” or “world” museums around. But these are really, at best, nationalist spaces. At worst, they are fascist ones.
There is nothing universal—in the sense of stateless, global or even cosmopolitan—involved in what you are up to.
Data can be stateless. I think we (the people) should build digital world museums that live up to the name. Citizens of the world cannot yet overcome national citizenship; I wonder whether data can be our avant-garde in that struggle. After all, it exists in territories that range from the somewhat uncontrollable to the outright unregulated and anarchic. Imperial museums are becoming vulnerable to hacks by decolonial minds. Datasets cannot be contained: Once online or on a public domain, you can add a copyright license but it means nothing. The virality and the afterlife of (cultural) data has real beauty and power. I like to call it, in the words of Sonia K. Katyal, technoheritage.{1} But stateless technoheritage does not mean random data without a context, or even without representation through classification. On the contrary, today we know that data without context is at best worthless and, at worst, harmful.
The conversation gets quite upsetting when museums use copyrights to restrict the distribution of knowledge and remixes derived from their physical and virtual holdings (how can a thousand-year-old artwork have a copyright, anyway?). Museums offer many arguments when trying to prohibit re-use in the public domain: They want to ‘preventcommercialization’ (except in their gift shops, of course), or even to protect someone or something from the “bad taste” of us peasant people and our shameless remixing 😉 I am not joking.{2}
In short, museum practices in the Global North seem highly anachronistic. The suggested ‘democratic potential’ of technology stands in sharp contrast with widespread institutional angst about their declining relevance, and their threatened status as gatekeepers of legitimate interpretation (Deutungshoheit) and representation. You are right to be afraid. Yeah, you’d better be afraid—because the digital is no slave to the original! And, believe it or not, many people do not even require ‘the real thing’ anymore.
In the same breath—one needs to respect (yes, R E S P E C T) the fact that not all cultural data wants to be free and dispersed. This is crucial. Datasets will only dance with you once you acknowledge, see, know, or truthfully relate to them. As a matter of fact, in very different cultures, on different continents, a dataset or 3D-printed object can hold the same power and enable the same connection as the original material item. This is the reason why some people think that certain objects should neither be exhibited and touched by custodians, nor digitized and their copies exhibited. These people view cultural data not just according to its inherent technological potential (as Western media theory so often does) but as something that is constantly being translated and given meaning by local versions of world-making, cosmologies, and ancestral knowledge. For example, in the case of the Tāmaki Paenga Hira Auckland War Memorial Museum, the Māori took the power to decide whether their objects, images, and ancestors should be digitized or not. The question is: How should the power to decide about digitization and thus representation be distributed? Who, in each case, should hold it?
Decolonizing databases is not just about single objects but entire knowledge systems.
One beautiful example is Mukurtu CMS (Mukurtu is “a Warumungu word meaning ‘dilly bag’ or a safekeeping place for sacred materials”).{3} It is an open-source platform and CMS (Content Management System) that serves diverse communities who want to manage and share their digital cultural heritage in their own way, on their own terms. Users include cultural protocols and traditional knowledge (‘TK’) labels such as secret/sacred, seasonal, or women-only.
Now with data-driven technologies like Artificial Intelligence (AI), the work of decolonizing becomes even more twisted… and fun. Fooling around with digital objects and releasing them from their proprietary museum systems and their commodity chains is an act of resistance against a pre-narrated archive. You gotta admit, any form of (techno)heritage is (data) fiction!
So, how will we use and shape technology to make sense of culture? Machine-learning makes certain patterns visible, including those that we didn’t know about—and those that are not talked about. In Wendy Chun’s essay ‘Queering Homophily’ she describes how ML can offer “the generative power of discomfort.”{4} Discomfort is indissociable from decolonization.
For emancipatory as well as for artistic practice a few things are worth spelling out. Most of these technologies are not yet mastered (nor fully understood) by the colonizer, that is, by the museum, this overheated, jerky colonial machine. The museum database is therefore still a political thing with emancipatory potential. One that is neither bound nor controlled by global power structures and which, on the contrary, can be a means of overcoming them. How exactly this works I will explain in a moment. But if AI is only as good as its database, the museum is probably only as good as its database, too. In a panel discussion at the World Economic Forum in Davos, in 2018, AI expert Jürgen Schmidhuber speculated that machines might soon be able to generate and gather their own data, no longer relying on open databases or permissions from museums or other institutions museums. This is the genesis of the (museum as a) database without walls.
In my recent work “Babylonian Vision” and “Neuronal Ancestral Sculptures Series”, realized with the help of ML experts Negar Foroutan and Melika Behjati, we scraped databases from the largest collections (all of them in the Global North) of Mesopotamian, Neo-Sumerian and Assyrian artifacts and got 10,000 images. We used these to train a neural network whose process was designed to yield abstract insights into the search for a visual language of form and pattern—delivering a speculative and anarchic archaeology.
A visual museum database contains images of original artifacts that can be used to train any Generative Adversarial Network (GAN). The input images of these databases are carrying time and memory themselves, such as patina or broken pieces. Of course, the AI doesn’t see a single image, but numbers only. This is a new form of image-making, with a latent space for new synthetic images that the neural network is opening up.
If you train the network with, for example, a hundred thousand human portraits, it will abstract the concept of a portrait and generate new portraits. The same goes for museum artifacts! GANs and their generated aesthetics certainly challenge our understanding of and obsessions with authenticity, originality and authorship. As Nora Khan wrote on our perception of generated images:
“our rational understanding of the process actually helps us enter how provocative this method of image creation is. And with this understanding, we can linger, analyzing, close-reading the image flow for symbols and meaning.”{5}
So, what’s the relation between image-making and GANs? Where can I see this secret latent space? GANs could be the “start of a new paradigm in making pictures”.{6} There’s no magic involved, only numbers. Nothing uncanny or dreamy to see here. And for this very reason, what we can grasp from the latent space is more fascinating—because it is the unexpected that we can find, subjective and interpretative, like an essence of many images, a mood without a frame, recalling the infinite. It reminds me of the Mesopotamian way of making images, as noted by Zainab Bahrani:
“The images I present are often infinite in their compositional form and conception. They resist the frame and are often depicted as segments taken out of the potentiality of an endless composition that can be repeated in an endless series of mirrorings, pulling the image into a vertiginous mise en abîme … For Mesopotamia, the place from which we have the earliest textual and archaeological evidence about concepts of the image and aesthetics, I make the case that images had a diachronic presence; they were seen as objects that transcend time and that carry or embody traces of time itself. ”{7}
The produced images can be even photorealistic. But they’re not photos, because each one is made up from the noise of hundreds of epochs (the generator’s repetition of learning, via the discriminator) by the neural network. How new and original are these images, really? Is this verisimilitude? And does this matter? There are even larger questions when it comes to what knowledge and forms in the arts are, as passed on through the generations and how we are inspired by them. For this reason, the output images are something that I refer to as “ancestral”: What we can see through GANs are the semantics of very large visual datasets.
And in the latent space, there is semantic content that has meaning for the human eye and mind, and this resonates. After all, what is at issue is the meaning that we give to data. It is impossible to control or foresee the latent space, because it is a space of the unknown and unseen, of an accumulation of visual epistemologies or knowledge systems. The abstraction of the output allows us to contemplate the images and their language as a form of visual memory not limited to the input objects—a memory that transcends them, and which is able to generate new memory objects; the potentiality of an archive of infinite abundance. Through GANs, one generates new materialities, which rise to the surface as the affective qualities of the original in a post-original form.
My contention is therefore that if ML is seen as a technology performing and processing our collective memory, it makes sense to apply it to big cultural data of the past, to generate and to give rise to original, synthetic images.
In the case of GANs there certainly is prevision, but humans can’t control or predict what’s produced. We have to let ourselves fall into the primordial nebula of our ancestors and of our circulating image-worlds…
Another important aspect that I tried to address in my “Babylonian Vision” work is the question of representation in datasets that are used to train AI. Today, datasets are trained with millions of images governed by large corporations: With ImageNet or Open Images, 80% of the content originates from the Global North. This is visual hegemony. While not being seen may often be a big advantage in the surveillance age, it can also lead to being dominated by Western visual culture.{8}
As of today, AI is a black box system that doesn’t show its workings and it is almost impossible to reverse its input data from its outputs. This can be harmful, because of privacy and bias issues. But it has turned into an advantage in the case of my project: Ironically, the chosen museums cannot make a legal case against me and this form of image creation, because they will never be able to prove which datasets were really used for training 😉 In this manner, the adversarial in GANs is the decolonial. In many instances, the black box is a problem, but in some cases, it can also be… well… magnificent, decolonial and emancipatory.
Let the data dance!
{1}Katyal, Sonia K. “Technoheritage.” Calif. L. Rev. 105 (2017): 1111.
{2}When students from Stanford University scanned Michelangelo’s David, they had to promise to “keep renderings and use of the data in good taste” because the artifacts “are the proud artistic patrimony of Italy” (ibid.,p. 1148).
{3}See the useful writeup of a podcast with a member of the Mukurtu team at https://theconversation.com/mukurtu-an-online-dilly-bag-for-keeping-indigenous-digital-archives-safe-112949
{4}Chun, Wendy Hui Kyong. “Queering homophily.” Zeitschrift für Medienwissenschaften 18 (2018): 131-148.
{5}Khan, Nora introduction in: Reas, Casey: Making Pictures with Generative Adversarial Networks, Anteism Books, 2019.
{6}Reas, Casey: Making Pictures with Generative Adversarial Networks, Anteism Books, 2019.
{7}Bahrani, Zainab: The Infinite Image: Art, Time and the Aesthetic Dimension in Antiquity, Reak-tion Books, 2014, p8.
{8}Shankar, Shreya, Yoni Halpern, Eric Breck, James Atwood, Jimbo Wilson, and D. Sculley. “No classification without representation: Assessing geodiversity issues in open data sets for the developing world.” arXiv preprint arXiv:1711.08536 (2017).