Irini Kalaitzidi: Feeding Birds

Photo: Stathis Doganis

“Feeding Birds” is a research project that attempts the poetic use of artificial intelligence (AI) in dance reading/viewing. The project involves the development of an archive of dance phrases being used as a dataset to train a Human Action Recognition (HAR) model. This technology, originally created to serve military surveillance practices, functions here as a poetic device that arranges human actions and dubs them with labels such as “feeding birds” or “folding napkins”. The dataset used to train the AI model is what will ultimately determine how the model itself will (mis)interpret the world, which label to assign to which action. Feeding it with the dataset is a process of shaping the model’s dancing memories and biases.

This research will lead to a performative act where humans and AI open a bodily and verbal conversation, like in a Telephone game where each side proposes a different reading of the same stage reality. A human dances and an AI model observes and assigns a label to the action it detects. The model returns this label back to the human as a choreographic score for his next action. A chain of successive translations: action-label-action-label and so on. In this work, the AI is invited to discuss the dancing body and trigger reflection on the ways in which we see and interpret dance.

Stathis Doganis

Creator's Note

“Feeding Birds” originated from my fascination and unease with the intersection between movement and classification technologies, and the human capacity to impose meaning through naming things. At its core, this research project critically examines how Artificial Intelligence (AI) technologies, specifically those designed for Human Action Recognition (HAR), clash with the fluid and elusive nature of human movement. Can an AI model, trained to identify a set of movements with predetermined labels, ever fully capture the nuances of an action? Can this action be fully and consistently contained within one single label, such as feeding birds, sitting, or riding a bike? “Feeding Birds” interrogates the rigidity of AI classification systems – predominantly used in surveillance and military contexts – against the backdrop of the inherent subjectivity and ambiguity of human movements.

A note to readers and my future self is that actions and movements are used interchangeably in this project, with the awareness that this is a soft handling of words that needs further investigation. For now, in my defense, this soft handing of words fits within the broader impossibility of using language in a rigid manner within the research’s overwhelming landscape of labels and annotations. As I elaborate later, this impossibility to accurately name a thing, process, or action became a core thematic thread of the project. But before diving into that –

The project’s inception was quite literal. While examining an existing dataset called Kinetics400 – a collection of labeled actions used to train HAR models – feeding birds was listed as one of those 400 foundational actions that an AI was programmed to learn about humans(!). I found that funny and bizarre, and so, the label stayed with me. When testing a model trained on Kinetics400, I quickly realized that without the presence of actual birds, the model could’nt recognize my reenactment of the respective action. The movement alone was not enough to be identified as such. This realization sparked the idea for “Feeding Birds” – a project that would explore how AI (mis)interprets and (mis)labels human movements.

To delve deeper, I decided to train my own HAR model on a custom dataset of movements that I collected from scratch. This process began in August 2023 during my residency at K3 Center for Choreography in Hamburg and culminated in September 2024 during my time at Onassis AiR. My pipeline included four cameras recording one performer from different perspectives at the same time, while they were repeatedly executing the same movement with small variations for 15 minutes straight. This resulted in one hour of recordings per action, and 52 actions in total, labeled with the numbers 1 through 52, to later train a HAR model on.

Following a gap between the first and the second part of my Onassis AiR residency, I came back in May 2024 with the intention to create a performance where the model would engage in a chain of translations and misinterpretations between movements and labels – similar to how a game of broken phone. In this process, humans would perform an action, the AI would suggest a label for it, and humans would then perform based on that label, followed by the AI generating a new label, and so on. For this to work, the dataset’s actions needed labels beyond mere numbers, so that the AI could associate specific movements with words.

The dance artists Maria Vourou and Veatriki Kapnisi and the sound artist Nikos Tsolis were invited in this process as performers and thinking bodies that would engage in dialogue with AI. We all soon realized that the most intriguing part of the project lied in the labeling of the actions – revealing our biases, aesthetic choices, and opinions. We were faced with this impossibility that I referred to earlier when it comes to naming: creating labels and categories that could adequately contain certain movements within them.

“There’s a kind of sorcery that goes into the creation of categories. To create a category or to name things is to divide an almost infinitely complex universe into separate phenomena. To impose order onto an undifferentiated mass, to ascribe phenomena to a category – that is, to name a thing – is in turn a means of reifying the existence of that category.”

Excavating AI: The Politics of Images in Machine Learning Training Sets By Kate Crawford and Trevor Paglen, 2019

“Feeding Birds” drew and keeps drawing inspiration from Kate Crawford and Trevor Paglen’s ideas on the sorcery of categorization – the way in which the act of naming or labeling imposes an arbitrary sense of order on the infinite complexity of the world. And the process of labeling actions in order to feed an AI model becomes a kind of modern-day alchemy happening behind closed doors. Veatriki, Maria, and Nikos have playfully embraced this concept, with the latter referring to us as a sorcerous committee for naming actions. Yet, beneath the whimsical nature of this metaphor lies a serious inquiry into the power dynamics and ethical implications of AI systems that define and categorize human actions, based on their heavily biased training dataset.

For Open Day #12 at Onassis AiR, the research behind “Feeding Birds” manifested through a performance that zooms into the act of observing and labeling movements to feed an AI model – an process carried out with unwarranted certainty, despite lacking any true epistemological grounding and being influenced by the annotators’ inherent biases. The performance mirrors the confidence of birdwatchers, who, armed with their own assumptions, trust that they can accurately identify every possible bird species they observe.

With the ongoing support of Onassis AiR and Stegi’s Outward Turn Program, a next iteration of the research, this time called “Occupied with Traffic”, was realized during my time in ATLAS at ImpulsTanz, in July-August 2024. In this solo performance, I further narrow the focus to a single action (from the 52 originally), this time challenging the audience to engage with the process of naming and categorizing in a more personal and introspective way. As the project evolves, whether under the title of “Feeding Birds”, “Occupied with Traffic”, or another name that may yet emerge from my admittedly love-hate relationship with naming, it will continue to explore the (im)possibilities of movement (mis)classification through AI, and by extension, the choreographic capacity of AI through (mis)labeling actions-movements.