Generative Representation

December 2018

Heather Dewey-Hagborg is a transdisciplinary artist and educator who is interested in art as research and critical practice. Her controversial biopolitical art practice includes the project Stranger Visions in which she created portrait sculptures from analyses of genetic material (hair, cigarette butts, chewed up gum) collected in public places.

Read full Bio

For the past six years I have been researching, writing and producing artwork engaging the methodology of "forensic DNA phenotyping". In brief, this is an emerging technology which attempts to depict a likeness of a person, to predict a person’s physical features - namely their face - based on DNA alone. It is truly a method of the present technological moment, bringing together research in data science, machine learning, 3D scanning, 3D modeling and genomics to produce hybrid predictive models. While the technology is not speculative, it is not accurate either.

Forensic DNA phenotyping has an air of infallibility to it because it is a form of DNA profiling technology, and this is easily conflated with DNA fingerprinting, the "gold standard" in police work. But phenotyping works differently. It is predictive, not identifying. It uses machine learning to make inferences about a person's traits based on their DNA, and like most machine learning models it has many problems with bias and stereotyping, a particularly dangerous territory when working with genetics in the field of policing, a topic I have written about extensively elsewhere.

In this essay, I would like to explore a different aspect of this technology, one that has not been widely discussed, which might begin with the question: is forensic DNA phenotyping a photographic process? The question is significant because photography generally implies a subject which can be represented; it seems to have a certain claim on ‘reality’. This is not an abstract or esoteric consideration, but one with real social and political consequences for how we see the world around us, and how we assess the truth claims of the images which circulate in our networks. It is of particular importance when considering the production of images used in police, state, and disciplinary contexts.

Generative representation is the production of images that appear ‘real’, although they are constructed using completely artificial means. It describes the way in which an algorithmically produced image, like a portrait, might feel like an index to a real world subject, even though it is entirely contrived. This phenomenon gives the generative image an authority that is essentially borrowed from the long history and tremendous power of representation in modernity.

Today, as we are engulfed by images of dubious origin, propagated through politically structured networks, curated by opaque and sometimes discriminatory algorithms, it is an important moment to reflect on the power of the image, and its increasingly computational nature. In approaching this question it is worth taking a minute to discuss how the phenotypic image is produced, and to compare this with what we think of as ‘photography’.

There are two primary forms of data that underpin the phenotypic model: genomic data, and 3D facial scans. One can intuit how these two datasets could be mined for correlations. For example, a data scientist might look for connections between the DNA code and the structure of the face that tend to co-occur; connections that don't necessarily imply causation, but rather, simply appear together more often than one might expect. And this is the fundamental insight that underpins the phenotypic impulse: the hard scientific problem of how genes actually create the proteins which make up human appearance doesn't have to be solved. Instead, it is ‘good enough’ to make a prediction based on generalisation from piles of data that can then be operationalised in contexts like policing.

The subject’s DNA profile is fed into an algorithmic model of a face which can be morphed along axes related to these correlations. Thus, behind the scenes, the code is essentially dragging a slider to make a face lighter or darker, more male or female, according to the parameters and limitations of the underlying model, and characterised by the data that determined the model in the first place.

Ultimately this results in a workflow in which genomic data is fed into software and a 3D face is generated as an output. Depending on the code of the software itself one might also be able to generate many possible permutations of a person's face based on the structure of the data. For instance, I have used this process in my project ‘Probably Chelsea’ where I algorithmically generated thirty different possible portraits of Chelsea Manning using an analysis of her DNA.

If a trait is associated with different probabilistic phenotypes (as they generally are), then a set of possible faces might better show the potential of the person’s DNA than a single image. So the output of the phenotyping process might be a single face or a slew of probable faces. There is no lens involved and there is no camera, but there is a production of a series of digital images.

Having unpacked DNA phenotyping, let’s return to the question of whether phenotyping is a form of photography. In exploring this question, I’d like to draw on Daniel Rubinstein’s essay 'What is 21st Century Photography?' commissioned by The Photographers’ Gallery, and the series of lectures published on his website

Drawing on Jean-François Lyotard, Rubinstein suggests that the intellectual project of modernity was to represent the world, and we saw this unfold in technology, in political systems, in art, and in economics from the 14th century onward. This, he reminds us, was the task of painting until photography displaced it as a primary representational form. In a time of science, capitalism, and democracy he argues “representation was everywhere.”

Rubinstein sees our contemporary moment as constituting a new technological and philosophical shift that calls representation itself into question. As daily life becomes increasingly abstracted, virtual, informatic and algorithmic, representation begins to lose its stability. We live in a world of images but the character of those images has morphed into an “immersive economy”. The image is a computational product, the outcome of an algorithmic process.

This demands a new approach—I would argue to every field—but specifically for the purposes of our discussion, to visual art and photography. An approach that utilises algorithmic methods to deconstruct our automated and virtualised society.

In Rubinstein’s words:

The problem is that in a post-Fordist society the locus of political agency and of cultural relevance has shifted from the object – as visually arresting as it might be – to the processes that (re)produce and distribute the object. Processes, however, by their own nature, are less visible and less representational than objects… In the 21st Century, photography is not a stale sight for sore eyes, but the inquiry into what makes something an image. As such, photography is the most essential task of art in the current time.

This perspective paves the way for a slew of new imaging practices that are deeply computational. It opens the field of photography to non-lens-based generative processes, and it calls for an art form that exposes algorithmic production. The phenotypic image then, might seem to be an almost ideal candidate for this demand of contemporary photography. But it inevitably leads to a deeper question about the nature of representation itself.

If phenotyping can be held up as a prime example of new algorithmic, post-representational photographic practice, is phenotyping non-representational? And if we follow this question further it brings us to yet another embedded question of whether a model is a representational, because genomics and phenotyping, along with most of our algorithmic surroundings, are all generated by machine learning models. To analyse the representational power of generative images, we must first understand the model from which they are produced. To understand the truth status of these pictures, we need to unpack their representational authority.

What is a model?
In a general sense a model is a reduced description of a complex real world system designed to fulfil a certain function. In machine learning, a model is an artifact produced through a training process in which a learning algorithm is exposed to data. The model is the outcome —the file or allocation of computer memory—which contains the structured result of a learning process.

For example, a model might be trained on input data that consist of images of fruit, tagged with the name of the variety of fruit. Once trained, the model will take any input image and assign it a category of fruit - regardless of whether the image has anything to do with fruit. Fruit determines the limits of the model’s universe, and beyond that the specific fruits it was exposed to define its inductive bias.

Does this model represent fruit?
We could see this model as representative, in the sense that it is sampling concrete images of fruit, this is the foundational basis of the model. Without a dataset of fruit experiences there would be no model. However, to see this then as a simply representational system strikes me as too simplistic.

The model is constituted by a subject that is not singular but multiple. It is always simultaneously all of the fruit it has ever seen, the statistical relations of this data, and the spaces in between actual data points. It is an attempt to capture an ‘essence’ of what constitutes a specific category of fruit, i.e. apple-ness, to move beyond singular representation of the subject and into the realm of the abstract. It is important to recognise that such a form is always contingent and always limited. It is not, and should make no claim to be a universal depiction of a fruit. It is precisely this leap from sampling and analysing data, to the claim of essence, from which so many problematics in machine learning derive. Nonetheless, the model contains latent potential to concoct images of fruit that have never existed; in a sense to imagine fruit that goes beyond the simple exemplars of its experience.

So the model, then, has some characteristics of representation but also transcends it. If a ‘sample’ is representative of a homogenous set, if a politician is representative of their constituency, and if photography as the capacity to frame and ‘represent’ a subject, a model appears to share this direct relationship while being structurally much more complex.

There is therefore a danger in seeing models as simply a new form of representation, because what they actually do, the way they function and act in society is fundamentally different. The model is shapeshifting. It can adopt a host of forms each of which may appear as certain, accurate, and as true as the last, but in reality the model is a constellation; a fluid space of possibility.

The phenotypic image shows us the power of what we might then call generative representation: the production of an image that feels real, that feels even like a direct representation, but is in actuality a phantom.

This is why a single image produced through the DNA phenotyping process is always a lie. It feels compelling, it resonates with the viewer because it speaks to our representational history, but this singular image hides the incredible complexity and combinatorial potential that underlies it.

Returning to our question, it seems that forensic DNA phenotyping may indeed be a form of post-representational photography, but it points to an area in which we should exercise caution, because the phenotype borrows authority from the representational paradigm in a manner which might mask its intrinsically generative nature.

This brings us back to Rubinstein’s call for an engaged 21st century photography, one that is not backward looking but staring the future in the eyes with a critical gaze, using tools which interrogate algorithmic forms and aesthetics. And here the phenotype hovers before us, as if a ghost from another era, exposing how easily we fall back into representational assumptions, and how sticky and complicated it might be to let go.