A way to rectify bias in AI
Jean-François Puget, a data professional who holds the title of IBM distinguished engineer, delivered a keynote presentation at the Strata Data Conference in London recently. The theme of his talk was the consideration of humans when bringing in automated processes. His main example focused the resistance people can have to new automated processes that work that they are not used to, but he also spoke of bias in artificial intelligence saying it “can be a bit more complex.”
Puget gave an example that he had worked on with his team, a study on facial recognition. They looked at members of parliament from all over the world and ran experiments with facial recognition services from major vendors. The found that accuracy of the services varied wildly according to the race and gender of the subject. He said: “We had a 0.3% error rate for white males. We can improve a bit but it is near perfect.” In contrast the accuracy of recognition of black women was “a disaster,” with a 46.5% error rate.
Puget said this was not down to the algorithm or the model. It was due to the training data not having enough diversity. The model had been trained with many white and male faces but so many black and females ones.
"Make sure you don't inject bias in your machine learning technology."
However, he and his team were able to fix it by retraining the model using a more representative training set. “This is just to show that the human factor here plays a role in how you train your models and what data you use.” He went on to say: “If there is bias in your data, the machine learning technology will learn it.” He ended with an instruction: “Make sure you don’t inject bias.”
This is probably easier said than done. However, Joy Buolamwini has made great strides in highlighting and addressing the problem. The research assistant at MIT Media Lab, who investigates social impact technology, said that many developers use off-the-shelf libraries when they are developing facial recognition systems. This is to save time and avoid duplicating efforts, but detrimentally this leads to many computer vision projects sharing the same code, and as such “any bias in the system propagates widely and plants a ‘coded gaze’,” she said.
"Computers view the world through a coded gaze."
Buolamwini said: “Computers view the world through a coded gaze. They digest pixels from a camera in dictated ways.” A lack of diversity in training sets meant they couldn’t detect faces like hers.
As a student at Georgia Institute of Technology and later as a graduate researcher at MIT Media Lab, she has worked on projects involving interaction with social robots, and the Aspire Mirror which projects digital masks on to reflections. The inability of the machines to recognise her face impeded her somewhat, so her solution was to borrow her roommate’s face or to code in a white mask.
Buolamwini gave a presentation on her experience of this ingrained problem in a Ted Talk, and as a solution has created the Algorithmic Justice League, a platform that allows people to do three things. Through this platform, people are encouraged to increase awareness of algorithmic bias by highlighting it through documentaries and other media. They can report their own experience of bias in a technical system. Finally they can request a bias check and to help ensure that the technology they are developing is more likely to work with and for a diverse group of users.
Making sure you don’t inject bias is no small task. The developer has to be aware of its existence and take proactive steps eliminate it or avoid it from affecting their system. Fortunately, there are now tools to make that task a little bit easier.