top of page

Contextual Normalcy

* Note - all work on this is done was done as part of the collective Feminist.AI. While we all participated in brainstorming, design ideation, and workshops, not all of what follows is my work. My particular role on the project as that of AI development and data lead.

01 Motivation

Mental health is a topic at the forefront of news stories and community conversations and efforts in schools, workplaces, and health systems. Many feel that current approaches fall short in addressing their specific needs that are crucial to improving their current situations and has given rise to the radical mental health movement.* Grassroots and community approaches are invaluable in improving mental health approaches, but we feel that a wholesale redesign of frameworks and approaches is key to improved outcomes for those in need of health care. This serves as our primary motivation in this work, as we seek to include all voices, from trans artists in LA to pilots in Alaska to Native American communities in Oklahoma.

CN1.png

02 Why "Contextual Normalcy"?

LHI.png

One of our main criticisms of current approaches to mental health is that models are based on an idea of universal feelings and human experiences. Based on the work of Sara Ahmed, we respect that feelings are not only individual, but social as well, and that while the physiological response, the emotion, is a human universal, the mental maps are based in community-centered contexts - where “community” is not referring to the work of “imagined communities” ala Benedict Anderson. We refer to communities here as groups of humans with values, interests, or identities in common, and contexts here can refer to everything from a shared hobby to languages to temporal contexts.

03 Findings from the participatory workshop

SYSTEM: Currently manifests as Mental Health

INPUTS: Crowdsourced data about our feelings and thoughts about our feelings. Location data. Varied voluntary demographic information. 

RULES: Supervised and unsupervised machine learning algorithms

OUTPUTS: Visualization, info for inputs, new words, new classifications, new language of emotions, contextual normalcy app, AR distributed emotion app, intelligent location based experiences  

CULTURE: We primarily focused on individual and collective cultures.

PERSPECTIVE: We prototyped with several perspectives from groups to individuals around the world.

UNWANTED BIAS REDUCTION: By identifying who designed the data collection, who contributed to the data, who created the models, which models where used, and why.

The findings from informed the Contextual Normalcy Project

04 Designing for feelings

Design so that the feeling can be interacted with
Design so that the feeling can be abstract
Design so that we don't know what the feeling looks like

Design so that the feeling can mutate in shape and form over time
Design so that the feeling can act as a data point for machine learning
Design for privacy as a First-class citizen of the system

remove feeling.png
remove feeling 2.png

We asked a select group of users to imagine how they would keep or remove an emotion, from within and outside of their bodies. We also asked the users how their bodies would emit certain feelings. These movements informed the behaviors given to the abstract feeling cells, along with the how we design the feeling release for the phone experience.

​

Through this approach, we are able to reinforce the importance of collecting interesting data insights from multiple and varied realities for the larger aim of expanding AI design. The information also helps us understand the cultural perspectives on the roles of our bodies, location, thinking about feelings, and even color to paint a larger picture of what it means to be an embodied human with a distinct cultural background.

05 Embodied data sorting

Embodied data sorting is the approach Feminist.AI used in XR to annotate emotional data by mapping it to virtual reality. By creating objects to screen and exporting them to 3D experience, we can design across realities. This allows us to sort cross-culturally – while one culture might think of an object as representing a feeling, another might think the object represents another and would place it in a different virtual location corresponding to the feeling.

ContextualNormalcy.png

In the summer of 2018, Feminist.AI members Carey Crooks at the Media Arts Practice at the University of Southern California and Justin Lee at UC Santa Cruz developed our first prototype of an AR data sorting and rating software. Embodied Data Sorting allows users to engage with their feelings as embodied avatars and to explore manifestations of how those feelings are expressed in the AR space. Our next step will be to expand the prototype, gather data, and improve iteratively on the software as we deploy it globally and then use the data in machine learning.

06 AI design

Underrepresented voices across cultures are used to identify patterns and thinking around behaviors and traits that we use to understand how we feel. Using community- sourced questions and data, we use these patterns to show traits of reported feelings across cultures. Using speculative, interactive, and AI design practices, we discuss current challenges and vision alternative futures for mental health diagnosis and treatment. In this research project, participants co-create new thinking about classification, diagnosis, and treatment and produce a critical making response to existing thinking and products.

​

genderidentity.png

female-identified

male-identified

Gender identity and happiness on a spectrum

SUPERVISED LEARNING

We used supervised learning to engage in classification for the first part of our research. We will pair the keywords from the questions with the primary feeling word in our question, and use this as our training data.

UNSUPERVISED LEARNING

One way to explore context without pre-assumed judgments is to used unsupervised learning techniques such as Latent Dirchelet Allocation (LDA). As all of our data thus far has been in English, we are able to use LDA to find different ‘norms’ within what we have - as we have found two different expressions of joy in American English.

NATURAL LANGUAGE PROCESSING

Two vital factors go into addressing potential bias - one is taking a language- agnostic approach to the data, and the second is maintaining an awareness of cross-cultural understandings of emotions and behaviors. Language-agnostic approaches to natural language processing is often a challenge due to text processing libraries existing only for national languages in countries with significant research and development budgets, such as English, French, German, and Japanese. To address these, we can focus on semantic approaches such as Word2Vec (there are pre-trained models for 30+ languages available), and annotation by native speakers judgments for classification if necessary. Cultural and social practices are not a technical problem, and require both a background and awareness of human behavior that is not limited to one’s own and demand input from a wider-range of collaborators.

Combining all of these using document similarity, we construct new normals - multiple “normals”, i.e. contextual normality.

document similarity.png

07 Current status

As of early 2020, Contextual Normalcy is on hiatus while Feminist.AI focuses on other projects.  I have left the collective to focus on other projects, but will likely revisit it with the other practitioners when we have time again post-pandemic.

bottom of page