The fibers of Maya’s eyebrow hairs rise together like a wave, creating small, shallow wrinkles that undulate upwards to the top of her forehead. Her eyes and lip corners follow as she takes in the image of my good friend, Alex. I recognize this facial ballet as her excitement, but watch as it is met with a monotone “Hi” from Alex, who quickly picks up his iPad and continues gaming. I feel a familiar pang of sadness and glance back at her confused face.
I replay the moment in my mind, trying to understand why I recognized Maya’s excitement as happiness while Alex registered anger, by his account. Her eyebrows rose entirely, not halfway, indicating happiness, not anger. Unfortunately, Alex cannot make this distinction because Alex is autistic.
Searching for a way to help him analyze faces, that winter I developed FaiCE AI, a visual recognition program that reads facial expressions to determine one of five core emotions: happiness, sadness, anger, fear, disgust. The program locates facial landmarks (eyebrows, nose, ears) and uses the distance between them to detect which of the five emotions is present, and then, in real time, notifies the user.
And yet, after weeks of developing and researching, the error rate of FaiCE AI was still too high. After several debugging sessions, I theorized that to make more accurate predictions I had to incorporate more than one type of machine learning model. I looked into different machine learning architectures and explored how to combine them in the most efficient way. I determined which parts of each model to implement by plotting training and validation curves, thus evaluating each candidate model. But even with the perfect mix, the code was right less than half the time. I had a hunch that I was overfitting the model to the training data.
And so, after simplifying the aggregate model, I felt a rush of excitement — the accuracy assessment revealed that my program had become more precise in identifying emotions in real time than the average human. This would undoubtedly help people with autism recognize facial expressions and I uploaded it to Github so that anyone could use it. Since we so often rely on reading human emotions in life, I want to continue devising ways to assist people with autism navigate social interactions and become more independent.
Creating FaiCE AI made me recognize how vital an emotional connection is to designing lasting solutions for others. My desire to help Alex motivated me to persevere amidst initial setbacks in developing a solution. I wanted to support him as much as he had supported me as a friend: when we moved from the Bay Area to Austin, Alex was the only person I knew in my new school as he had made the same journey three years prior and he made me feel at home. Although the production of FaiCE AI required me to study large datasets and analyze multiple algorithms, I never saw my work as simply making sense of a bunch of numbers, it was about Alex.
Often, when we think of global change, we picture helping those far away, in grand visions of making an impact for the world to admire. However, what I realize now is that changing the world does start with helping those around you, those who you love and care for.
As I am about to go to college and connect with a new group of people, I want to continue to understand issues and look for solutions which yes - will flex the muscles at the sides of my mouth, indicating happiness.