If an AI could achieve the level of complexity required for mental states, could it also undergo mental illness?

Between 2017-2020 I worked on an excessive hands-on research project titled Marrow, which aims to speculate about the possibility of mental illness in machine learning. A project that has developed since to an immersive theater experience in collaboration with Avner Peled. During the dvelopment we developed and released creative tools: Small dataset preperation, collaborative tool when working on GAN latent space and generating images in real time using 360 camera and GuaGAN. 

Marrow is an immersive, interactive installation, exploring the possibilities of mental disorders in machine learning models.

Grounded in hands-on research and speculative design, the participants experience the story of four AI models desperately trying to function as a family having dinner. By playing the role of each family member, the participants embody four different learning models, each with their bugs and glitches leading to peculiar behaviours and an overall dysfunctional family.
The project allows the paticipant to get a unique and personal glimpse of how AIs might integrate into our lives if we were to extrapolate the current learning models into fully-fledged thinking entities. By telling the story of a broken family, we learn not only about the flawed nature of these models but also about the problems of human thinking that developed them. 

In September 2019, we started the development phase under the co-production of the National Film Board of Canada and AtlasV.
With: Philippe Lambert (sound design), Paloma Dawkins (animations).

This allowed us to revisit a few of the decisions we’ve made a year earlier and question whether they are aligned with the project’s goal and calls for action.

Marrow highlights glitches found in machine learning models to reflect on the possibility of mental disorder in AI, and on the social constructs that lead to those glitches, offering new paths for social thinking.

During the Marrow development phase, our goal was to understand how to work with the results of a distorted family, adding real-time interactions with the users and creating more interventions with the GAN model. We looked for ways to link StyleGAN to the story, along with other models and visuals, wanting to raise a conversation about distorted outputs, and how they highlight the confusion when one’s own sense of expression slips away.

My role included: creator - project lead, script-writing, storyboarding, dataset research and scraping, Unity animations, Touchdesigner interactions, user research, user experience, programming.

While working on Marrow we train and build our on tools and dataset.

STEP 1. training on Perfect Families Dinner:

we wrote about the process of training on small dataset here.

Step 2. Developing tools to support the experience and the creation process.

we wrote about our Latent sapce explorer tool, and about the way we used GauGAN. 

Step 3. Installation
Documentation from the installation at the NFB lab, Montreal January 2020

︎   ︎