1. Course description

2. First Ideas

3. Final Idea

3.1. Motivation

The motivations to take this idea as our final project came from our wish to show the public what AI (artificial intelligence) and ML (machine learning) are in a critical way. Since we started the course and confronted ourself with the topic we surprisingly realized how much data do system and products gather from us. The topic of biases in machine learning also call our attention since machines learn from repetition and the information that they are fed with come from human knowledge, how does this biases could physically look?

Being confronted with a topic that is rather difficult motivated us to show and make people understand what is the connection between AI and ML in their data, their internet behavior and data consumption. Because the content that we all consume can form the opinion and perception of our surroundings; shaping our appearance to the Internet and the people that have access to it.

3.2. Research

Having our motivation we then decided to do some research on the topic and start thinking in which direction we wanted to go. How can our project possibly be visualized.

For the research, we read multiple articles about art and AI/ML, about chatbots, how do algorithms and stereotypes relate, etc. Here you find a list of the articles that we read:

• A ‘NEUROGRAPHER’ PUTS THE ART IN ARTIFICIAL INTELLIGENCE

• The Relationship Between Art and AI

• Welcher Chatbot ist schlimmer: Der rassistische oder der politisch korrekte?

• Stanford researchers use machine-learning algorithm to measure changes in gender, ethnic bias in U.S.

• Just like humans, artificial intelligence can be sexist and racist

• GloVe: Global Vectors for Word Representatio

• My family's Google searching got us a visit from counterterrorism police

• You’ll Shudder When You See What Google Knows About Your Web Searches

• There Is No Tech Solution to Deepfakes

• Machine Learning En Plein Air: Building accessible tools for artists

• Let Your Imagination Run Wild With This AI-Powered Text-To-Image Generator

3.2. Concept

Having our research done, we realized that we wanted to create a plug-in which would transform the profile picture of the user reading an article into a stereotype that the algorithm of the machine learning system was trained with. This stereotype should be connected with main keywords by an algorithm and “interpreted” by a Machine Learning System. The system is full with biases and our language stereotypes. Since the algorithm has no power of its own to “judge” and to make its own opinion – leaving away human bias – it will transform the picture of the person into a stereotype.

In order to achieve this, we wanted to create a database with pictures which represent common stereotypes in society that are related to some keywords. So if a person is reading an article, the algorithm will assume that the person is a certain stereotype and therefore merge the users profile picture with what we commonly connect – and can find in google image search – with the stereotype that would read that certain text.

This project should make people aware not only of how we judge and create stereotypes in society – as well as how the internet has been „trained“ to look a certain way – but especially on how machines learn from our behavior; what we read, write or what we say and internet and people who control it do not know who you are, but can misinterpret your actions online.

4. Mirror

4.2. Structure

The News platform contains an algorithm which is interpreting your profile picture based on our data set. Our dataset shows the relationship between articles with click bait headlines and pictures of stereotypes. At the beginning, the user gets informed by our terms of use which explain the transparency of our data collection. During the whole process of exploring and reading articles, our algorithm starts to mash up the users profile picture with stereotypical faces based on your browser history. For keeping the process transparent, a data visualization of the way how the algorithm works will be shown. This kind of interaction should clarify how we consume and create data and how machines can interpret it. Therefore it helps to encourage the user to reflect on the whole process by feeding the algorithm with information about alleged interests.

4.2.1. Stereotypes

Stereotypes help humans to categorize their surroundings and to define their identity through differentiation and similarities. Therefore we are creating Datasets for algorithms based on human biases. For our platform, we tried to define most familiar stereotypes according to the western culture. At the end we decided to reduce our examples on following terms with common perceived features:

4.2.2. The articles

Our articles were taken from different platforms like: vice, vox, huffpost and medium. Our intention was to provide articles that had clickbait headlines because this kind of language ensures accessibility for everyone. But it´s obvious it won´t be appealing for all target groups. For our prototype, we only focused on articles which contain certain buzzwords which could be interpreted as an interesting field of a stereotype. Therefore our articles are linked to characteristics of stereotypes based on explicit buzzwords.

The taught process that we follow to choose the articles was:

• What would the stereotype read?
• What would de developer feed to the machine as an article that the sereotypepe would read?
• What would the machine relate to a stereotype (creating it´s own biases)?

The articles:

Hipster/Influencer

1. https://www.vice.com/en_uk/article/jm5nvp/ruby-tandoh-eat-clean-wellness
2. https://www.vox.com/the-goods/2019/1/3/18159301/bathrooms-for-instagram-luxury-bathrooms
3. https://i-d.vice.com/en_us/article/43bn9b/if-youre-into-extravagant-vintage-stacey-nishimoto-los-angeles

Criminal

1. https://www.huffingtonpost.com/entry/police-mistakes-rape-murder_us_58a23d14e4b0ab2d2b184db9
2. https://www.huffingtonpost.com/entry/el-chapo-trial-begins_us_5beae722e4b0caeec2be4c02
3. https://www.vice.com/en_uk/article/bjxkbz/inside-the-one-uk-lab-testing-whats-really-in-your-drugs-safe-sesh

Unemployed

1. https://www.huffingtonpost.com/entry/being-evicted-from-your-home_us_5c1d33c8e4b05c88b6f885a9
2. https://www.vice.com/en_uk/article/5ga7pb/a-to-z-of-being-unemployed
3. https://www.huffpost.com/entry/stand-mixer-vacuum-coffee-maker-sale-walmart_n_5c26742ae4b0407e90826bee

Nerd

1. https://www.vice.com/en_us/article/59jn9x/i-watched-every-marvel-movie-in-a-week-and-now-i-love-this-bullshit
2. https://medium.com/@bobwu/the-20-most-influential-asian-american-leaders-in-technology-start-ups-8435f9f4676c
3. https://medium.com/mammon-machine-zeal/you-cant-go-home-again-the-legacy-of-the-angry-video-game-nerd-54d3fd6ed5ae

Nazi

1. https://www.huffingtonpost.com/entry/hungary-viktor-orban-muslim-invaders_us_5a53b6a3e4b003133ecb014c
2. https://www.huffingtonpost.com/entry/milk-white-supremacy-racism_us_5bffad35e4b0864f4f6a3e28
3. https://www.huffingtonpost.com/entry/daily-stormer-nazi-style-guide_us_5a2ece19e4b0ce3b344492f2�

Depressed

1. https://www.huffingtonpost.com/entry/mental-health-day-work_us_5c023517e4b04fb21168335e
2. https://www.huffpost.com/entry/read-this-if-you-take-anti-anxiety-medications-regularly_n_5be43812e4b0769d24ca123d�
3. https://www.huffingtonpost.com/entry/jk-rowling-depression-message_us_5ab10126e4b0697dfe1a88e9�

Gay

1. https://www.huffpost.com/entry/best-winter-moisturizers-for-every-skin-type_n_5c23e324e4b08aaf7a8d4c81�
2. https://www.vox.com/the-goods/2019/1/7/18172611/golden-globes-fashion-billy-porter-cody-fern-judy-greer
3. https://www.nytimes.com/2018/11/30/business/grindr-president-same-sex-marriage.html?rref=collection%2Ftimestopic%2FSame-Sex%20Marriage%2C%20Civil%20Unions%2C%20and%20Domestic%20Partnerships&action=click&contentCollection=timestopics&region=stream&module=stream_unit&version=latest&contentPlacement=3&pgtype=collection

Feminist

1. https://www.huffingtonpost.com/entry/woman-of-color-in-tech_us_5bf44547e4b0a7286a0e3992
2. https://www.nbcnews.com/news/us-news/michelle-obama-keep-fighting-gender-equality-even-if-it-makes-n918921
3. https://www.vox.com/2018/3/20/16955588/feminism-waves-explained-first-second-third-fourth

Shopalcoholic

1. https://www.huffingtonpost.com/entry/golden-globes-best-dressed-2019_us_5c32427ee4b0bcb4c25cc665
2. https://www.vox.com/the-goods/2018/12/4/18126188/morganite-engagement-ring-gemstone-diamond-alternative
3. https://www.vox.com/the-goods/2018/11/29/18106728/no-makeup-natural-paradox-branding

White-Trash

1. https://www.huffingtonpost.com/entry/white-woman-calls-911-over-anti-violence-protesters_us_5bc7ac17e4b0d38b5874adc1
2. https://www.huffingtonpost.com/entry/andrew-napolitano-trump-border-wall-emergency_us_5c344ba1e4b05d4e96bb9277
3. https://www.buzzfeednews.com/article/juliareinstein/this-pregnant-teen-wont-be-allowed-to-walk-at-her-christian

Sex addict/ pervert

1. https://www.vice.com/en_nz/article/bm9w7v/why-are-so-many-women-searching-for-ultra-violent-porn
2. https://www.huffpost.com/entry/masturbation-how-to-talk-to-kids_n_5c3020e1e4b073352832e3c6
3. https://www.vox.com/2018/12/4/18124120/tumblr-porn-adult-content-ban-user-backlash

Terrorist

1. http://www.inss.org.il/publication/isis-success-in-iraq-and-syria-strategic-ramifications/
2. https://www.huffingtonpost.com/entry/jenan-ayesh-anti-muslim-attack-dallas_us_5c34bc47e4b01e2d51f797d5
3. https://www.buzzfeednews.com/article/aliciamelvillesmith/explosions-reported-at-brussels-airport

4.2.3. Information Architecture

4.3. Machine Learning

To visualize our concept we choose to work with a Machine Learning System called GAN’s. GAN’S systems are able to use the derivative knowledge from training data for other interpretations. With the help of our ML system we wanted to evaluate facial characteristics of our stereotypes and project them to our users profile picture.

4.3.1. About

Therefore we trained our choosen system with all datasets to get an average imagery of every stereotype. The next step was to map those stereotypical Avatars on potential user faces. This was done by another GAN’S system. The final outcome were short clips which showed the metamorphoses of the potential user and its stereotype.

• Results average stereotypes using StarGan: When we started with creating our average stereotypes we had to realize our ML system was trained to only evaluate average features of faces and ignore those which are rarer or uncomment. The average faces therefore didn’t show the most prominent and stand out features we needed. For example tattoos, glasses or piercings were not recognized by the ML but still important features for some of our stereotypes.

• Results morphing Profile Picture using FaceMorphing: The machine learning of the average faces mostly erased all of the outstanding features we needed to be associated with each stereotype. So we decided to choose random images of our datasets for the metamorphoses. The ML generated morphing clips of random images of a stereotype and a potential user of Mirror. Stills of the clips we used as material for the evolving Profile pictures on our prototype as well as in the gallery.

In conclusion, we all agreed that for a more strict way of visualizing our concept we would have used the average faces. But considering the time we chose to used a serendipity approach. Especially to achieve more artistic and disturbing results it would have been great to write our own ML system.

Here the Github links:

• StarGan: https://github.com/alyssaq/face_morpher
• FaceMorphing: https://github.com/kubricio/face-morphing

4.3.2 The different algorithms

To connect the articles with our stereotypes our concept also included an algorithm. The algorithm recognizes all articles which are read by any user and connects them to a stereotype. Because we couldn’t train an algorithm we minimized its work by choosing buzzwords of each article. In consideration of the algorithm we defined a dataset of stereotypical characteristics and buzzwords. We then connected all characteristics of the stereotypes to all fitting buzzwords and characteristics of each stereotype. The algorithm is now classifying all articles trough buzzwords to its matching stereotype. For an overview, we created a small data vis which visualizes all buzzwords of the articles and their connection to the stereotypes.

4.3.3. Results

Here are some examples of the results that we gathered:

Average Results:

Morphing Results:

4.4. Design

After having the results of the machine learning system, we started designing Mirror.

4.4.1. Moodboard

We all had one vision in our head, having neon colors, dark looking and deformed images. Showing the colors of computer like components together with the distortion of reality.

4.4.2. Scribbles

Before starting to define how should our website look we did sketches about how should it be build, where should the information go, how does the storytelling works in the website and how does the information could be structured.

4.4.3. User Flow

Regarding to our concept we created a user flow chart.

4.4.4. Prototype Video

4.4.5. Click Prototype

you can find our click prototype under this link: Mirror

5. Conclusion & Learnings