7 min read

Face Space

Face Space

Welcome back to Socos Academy with our weekly updates on mad science projects, upcoming events, tours through the latest research, and, as always, a sample from one my draft book chapters.

This week: Faces!

Mad Science Solves…

I’ve been working on the latest draft of “SexyFace”, from Professional Mad Scientist, (more below). It is all about one use for facial recognition technology no one would questions…and a few others leave me feeling in need of a long, aggressive shower. While I was building the system beyond “SexyFace” in 2009, I ran a few additional experiments on a pre-existing database of faces (FERRET) that. at the time, was used extensively to build and assess face recognition systems. I used it myself a few times, first to train a system for lie detection and then to pre-train the model used in SexyFace.

This AI model was much more straightforward than them proto-deep neural network of SexyFace–I simply asked it to “learn the pattern” of the faces in the database. It pulled thousands of numbers for each face and teased out their minute similarities and differences. We were able to represent these learned variables in terms of red-green-blue pixel positions and interpret their patterns on an x-y axis. (See the figure at the top.)

The results were pretty transparent. Look at the faces on the left: from bottom to top, you see a transition from traditionally masculine to traditionally feminine features. Notice that the faces towards the top appear to be smiling more.

So, the data set is saying that if you want to be pretty and viewed as “feminine,” smile more. If you’re Asian, you are just sort of generically, ungendered Asian–there is no boy-girl thing for you to worry about. If you want people to see you as more than the color of your skin, just be white (pretty easy, right?).

I’ll be clear: going in, this model understood nothing about faces. Without any guidance whatsoever it was able to pull out patterns that demonstrated a tremendous amount of bias that existed in the database. Machine learning systems were supposed to use this database to learn a “language of faces”, but it turned out that the database had very few words for “Asian male”, and even fewer words for “black.” The palette was unambiguously poor in its representation of real people.

In our SexyFace story, we used AI as a diagnostic or causal tool to understand certain patterns across faces chosen by the player. In this little interlude project, we instead used AI as a tool for analysis, as a means of helping us understand what the data itself was saying.

Unfortunately, what it was saying was very problematic. It’s not that the AI or the data used in this project were racist; all our model did was learn what the photos in the dataset were already saying. It reflected gender and racial biases that existed in a supposedly unbiased dataset, and in doing so it reflected our own biases back to us. This project shows that the value of AI isn’t to help us combat bias, but rather to help us to better understand it such that we can begin combating it ourselves.

Stage & Screen

Ever want to pick the brain of an accomplished neuroscientist who's an expert in Artificial Intelligence and all manner of technologies?

Join our next Mad Science Solves, Ask Me Anything with Dr. Vivienne Ming, next week, Tuesday, July 25!

This event is a free talk series that delves into the messiest of human problems. Led by theoretical neuroscientist, entrepreneur, and professional mad scientist Dr. Vivienne Ming, these discussions explore some of the biggest challenges facing our society and consider how technology and innovation affect the choices we make. Complete with dirty jokes and sci-fi references aplenty, the informal structure of “Mad Science Solves” is designed to invite open participation as we aim to tackle the wonkiest angles of some of the world’s most complex problem spaces.

Next up! Vivienne will be speaking about the future of AI for an tech gathering in Indianapolis, IN this August.

Looking towards the Fall, Vivienne will be back on the move making stops so far in Berlin, New Orleans, Toronto, New York, Chicago, Athens, London…and more to come! <<If you are interested in pursuing an opportunity with Vivienne in or around these locations, please reach out ASAP!>>

Research Roundup

Facial Impressions Are Predicted by the Structure of Group Stereotypes

In machine learning research I did years ago, smiling and youthfulness strongly correlated with femininity in Facebook profile photos—and (weirdly), glasses-wearing with East Asian-ness.

A recent paper confirms what my work suggested, “people’s impressions of others’ faces are driven not only by [face shape and expressions] but also by learned stereotypes about social groups.” The title of the paper says it all, “Facial Impressions Are Predicted by the Structure of Group Stereotypes.”

Condemn the angry

Point 1,324,110 in favor of intellectual humility: average people and “professionals such as fraud investigators and auditors see ‘suspects’ angry responses to accusations as cues of guilt” when in fact, “accused individuals...were angrier when they are falsely relative to accurately accused.”

Combining the paper’s title, “Anger Damns the Innocent”, with the above finding that our very perception of facial expressions is biased by our out-group stereotypes, we might take it further: stereotypes damn the innocent.


Here is a teaser from the latest version of “SexyFace” from the upcoming Professional Mad Scientist.

learning and groundbreaking cognitive models to do what any aspiring, adventurous, heroic, world-disrupting Silicon Valley entrepreneur could hope to do: create an incredibly sleazy game for the horny masses. It was the heyday of tech–in the age of Zynga and Farmville, everyone’s goal was to grab as much money as quickly and cheaply as possible. Armed with PhDs in theoretical neuroscience and psychology, I spent all of my time in front of my computer perfecting the algorithm that was going to power a messy kind of love. While the other brilliant minds in Silicon Valley were busily growth-hacking high-burn social startups to disintermediate people, I was doing something different, something that could actually have a lasting impact on the world: I was helping people get laid. The game was called SexyFace, and the premise was simple. For free, we would use the cutting edge of artificial intelligence to find everyone on Facebook that you thought was sexy. And for five dollars, we’d find everyone that thought you were sexy.

While I was clearly making all the right choices in life, an orphaned nine-year-old girl was fleeing Afghanistan and crossing into Pakistan in a caravan of refugees. Dizzy from dehydration, knees weak with exhaustion, she focused on reaching wherever the caravan would take her. She knew nothing of the camp they were headed towards, and when she finally arrived, the line to enter continued for miles. The sun beat down from above, raising temperatures to 100 degrees Fahrenheit. There was no shade and the smell of human waste hung in the air, burning her nostrils. When she reached the front of the line, she looked up to see a strange-looking man staring into her face. She recoiled, but he passed her a cup of water. She gulped it down eagerly, the water streaking the dust across her mouth and chin. The man knelt down next to her and lifted a large camera to his eye. Terrified, the little girl gripped the cup to her chest as the photographer snapped a shot of her face, exhausted, dirt-stained, and lost.

Unaccompanied and with no family to protect her, she worked in the camp, a lost orphan carrying water, cleaning tents, and scrubbing clothes. The unrelenting stress of these early years became toxic to her body and her mind. She suffered from depression and PTSD, but there was no trauma counseling in the camp and no medical treatment. Later, the refugee leaders in the camp forced her into a “sale” marriage to a man who was cruel and violent. The accumulated stress and poverty wore away her cognitive control and emotion regulation. She had three children before she was sixteen, and just as her husband beat her, she beat her children. By twenty, she had had no formal education, a life spent entirely in a camp. Only one face in a sea of lost children, she envisioned killing herself every day.

If only her customer lifetime value were higher a VC might have funded an app to democratize her education. But in other news, SexyFace worked astonishingly well. The game would show a bunch of faces, male and female, young and old, from around the world, and ask players, “Who’s sexy?” With each selection, the AI driving the game would probe the borders of the player’s kinks, until after just a few rounds, it had it down. You didn’t even need to know what “sexy” meant to you in order to play–the damn game would figure it out! South Asian male with muttonchops and a sense of ennui? As long as you were consistent in your judgments and there were enough positive examples in the dataset (i.e., everything we could pull off of Flickr and Facebook) it would find all the other faces that tickled your fancy (and maybe tickled a whole lot more). Got a Costanza-esque thing for a pinkish hue? SexyFace has your pinkish dream. Does competence in a partner drop your knickers? SexyFace has you covered. (Got a foot thing...well, we did think about doing SexyFoot at one time, but I kind of suspected that the necessary training data couldn’t be found. I was horrifically wrong and can never unknow that truth.) Whatever your kink, Sexyface would find that pinkish, South Asian foot with a sense of ennui for you...

Subscribers can read the entire chapter here.