Fighting Algorithmic Bias in A.I.
“A.I. software is only as smart as the data used to train it. If there are many more white men than black women in the system, it will be worse at identifying the black women.”
Last week the New York Times highlighted a study by Rhodes Scholar and MIT Media Lab graduate researcher Joy Buolamwini that demonstrated the inherent bias in three major facial recognition providers.
At Smile Identity we’re well aware of this problem and are actively working to solve it for our specific use case — using one’s face to prove your identity when signing up for a financial account or approving a high value transaction.
Unfortunately, this problem is not new. Many technologies have been created with built-in bias against people of color. My co-founder has been aware of this problem since he worked on the old NTSC video standards in the 1980s. The standard made it harder to get quality video of darker skin. Before being adopted, it was either not tested on a truly representative sample of humans or the officials who signed off on it didn’t care that people of color were not shown in the best light.
More recently, Ava Berkofsky, director of photography for the HBO series “Insecure,” discussed this issue in an interview about her work on the show.
“When I was in film school, no one ever talked about lighting non-white people. There are all these general rules about lighting people of color, like throw green light or amber light at them. It’s weird.”
When we started Smile Identity we knew that most of the face recognition algorithms that were available had been trained on white males, and that relatively few people had gone about trying to rectify this.
While we chose different algorithms from the ones tested by Joy in her study we also found that individual off-the-shelf algorithms alone were not good enough in their current form to achieve the accuracy levels required for financial services use cases in Africa. They had all been trained on biased datasets. A biased dataset results in a biased set of outcomes. If you train an algorithm on a euro-centric population that algorithm will often get confused and drop its confidence scores when it encounters non-euro-centric users. Sometimes, it won’t even recognize a face.
To handle this, we put together an ensemble of algorithms and some user interaction techniques that allow us to get improved results over time. The algorithms compete against each other with the results reviewed by humans. As we recognize correct or incorrect outcomes we train the machines to filter and self-adjust based on image type and quality. Since our initial focus is Sub-Saharan Africa, our dataset is majority of people of color. This gives us a unique advantage in catching bias in algorithms — imprecision on dark faces — early on.
As new algorithms and best practices become available we’ll weave these tools into our system and continually test with the goal of creating a neural network that is more accurate on darker faces than any generic off the shelf facial recognition algorithm.
At Smile Identity we’re committed to building a standard that is inclusive.
We are grateful for the chance to try and right one of the technical wrongs in a field like artificial intelligence that will be incredibly important to people all over the world in the future.
Below is Joy’s Ted talk on the importance of fighting bais in AI.
In her words: “You can’t have ethical A.I. that’s not inclusive, and whoever is creating the technology is setting the standards.”
