Who Are I (Master Thesis)


Research / Design / Writing




Thesis Studio


Nowadays human generated data is an economic resource, a playground for quantitative social network analysis, classification and algorithmic profiling. Individuals today are no longer defined only by their declaration of self-identification, they are also a representation of their data, interpreted and transformed into proprietary bodies of Measurable Types which have their own histories, rationales, and politics, detached from individual, social contexts.

Who Are I is an installation that transforms my personal user classifications, as defined by Google’s Ads Personalization, into a generative self-portrait that prompts reflection on the genesis of algorithmic identities in today’s networked societies. In a large-scale video installation, I'm presenting fragmented portraits of individuals that have never existed – faces generated through Artificial Neural Networks – that form disembodied visual metaphors for networked systems of identity formation applied to all aspects of digital life.

Each portrait fragment represents a single vector within a complex, multi-dimensional, identity-determining latent space. A recognizable portrait is only visible when viewing the accumulation of all pieces: it is all of them at once and yet none of them. Always somewhere in the space in-between, stuck in dimensions that cannot be comprehended and upon which there is no influence or control, the figures depicted perpetually shift in a constant, mutative flow of data.

View full paper


Behind Who Are I lies a technology called StyleGAN [Style Generative Adversarial Network] which is capable of generating photorealistic images on the basis of training data. StyleGAN is built upon ordinary Generative Adversarial Networks [GANs], which are Artificial Neural Networks [ANNs] in the field of Generative Models. Their peculiarity is to learn the probability distribution of training data and subsequently generate new samples from that distribution.

A GAN always consists of two basic components, a Generator, whose task it is to synthesize new samples, and the Discriminator, which takes samples from the training data as well as from the output of the Generator and evaluates whether the latter are real or fake. To generate new samples, the Generator needs an input in the form of a random vector from latent space, which at first consists of pure noise. Over time, however, the Generator learns from the feedback of the Discriminator and transforms that white noise into more realistic images. The discriminator also improves over time by constantly querying and evaluating real and generated samples.

Dataset Creation

In order to synthesize potential visual manifestations of my algorithmic identity, I had to compile a dataset of images that were based on my personal classifications. I took advantage of the interpretative classification logic built into Google’s search engine and searched for images that matched my user profile, looking for combinations such as male+german, male+design or sport+male.

Therefore, I developed an image scraper in Python that allowed me to search and download images from Google in an automated fashion. The scraped images were not only portraits, but also full-body shots of people and images of faces at various angles. As I only needed portraits, I used an open-source Python library called autocrop, which is based on OpenCV’s face recognition software, to crop the images down to their faces. The final result of this laborious process was a directory structure with hundreds of portraits, representative of my algorithmic identity.

Training StyleGAN

For the training of my StyleGAN model and the creation of the synthetic portraits, I used RunwayML, a tool which introduces an intuitive visual interface to Machine Learning. For its training processes, RunwayML is currently using Transfer Learning which allows users to upload their own datasets and retrain the StyleGAN algorithm in the likeness of their own data.

RunwayML’s cloud-based architecture enabled me to outsource all computing-intensive processes and helped me to achieve good results rapidly and with little to no infrastructural efforts or costs. All of the above made RunwayML an ideal tool for the creation of my synthetic portraits that now only had to be generated.

Installation Design

Each portrait that I created constitutes a fragment, the sum of which composes an overall picture of my algorithmic identity. The idea of the fragmentation of an overarching body became the conceptual motif for the formal development of my installation.

I began developing my installation by means of a grid, subdividing the generated square portraits into individual surfaces of different dimensions. Each surface depicts a certain section of the human face [eye, mouth, nose, etc.] and allows for a focused and detailed examination of the respective body part. This effect is further enhanced by the explosion of the individual surfaces into space; the distance between each surface measures 50cm.

Their spatial arrangement, however, results in a distortion due to the depth perception of the human eye, which causes surfaces further in the background to be perceived smaller. By means of the Angular Diameter, a concept derived from astronomy, I was able to calculate a scaling factor [+12.5% / plane] that compensated for the spatial distortion, and thus again created a uniform overall picture.

Final Touches

Projectors are used to project sequential animations onto the individual surfaces. These animations are looping through different portraits, showing alternate versions of the same facial fragment [eye A, eye B, eye C, and so on]. Together, the individual fragments create ever new combinations of face compositions. The resulting collages are visual metaphors of my personal, identity-creating latent space, and thus self-portraits of my algorithmic identity.

A collection of my user classifications lie on the floor as extruded three-dimensional letters arranged in the format of a JSON file, an open standard file format that uses human-readable text to store and transmit data. This data visualization creates context and serves as a visual link between the digital and the physical – my user classifications and their metaphors.

Explore more work