Federal Study Finds Racial And Gender Bias In Facial Recognition Algorithms | NBC News NOW

Channel: NBC News
Published: 12/21/2019 05:44 AM

A new study from the Department of Commerce says that facial recognition technology may have more trouble recognizing people of color and women. NBC News’ investigates the potential downfalls of this new technology. » Subscribe to NBC News: http://nbcnews.to/SubscribeToNBC » Watch more NBC video...

[ music ] today we're in our mixed reality, studio to get the big picture on facial recognition. Okay, so when an artificially intelligent facial recognition program sees a human face, like this tell us what it's looking for first well, there are lots of methodologies that we can look at one in particular called open face. The first thing it needs to do is find out whether there's even a face in t ...
e photo, and it does that by looking for 68 different facial landmarks that typically represent what we would think of as a visibly human face within those 68 different groupings. We'Re looking for edges or gradients and the edges typically represent a group of pixels, that's the samecolor or that shifts from a darker color to a lighter one in the open face paper, there's particularly eight different landmarks that represent a human eye and many others comprise an Eyebrow or a mouth or nose anything along those lines that you might see when you look in the mirror. Well, now that we know we have a human face, the goal is to get it into a format. That'S uniformed and using these 68 points as anchors, we can then scale or rotate or even adjust the angle of the individuals face. So this is gon na allow the program to be able to make it bigger or rotate it a couple degrees and still be able. Toeffectively understand what it's looking at exactly. We want every face that we're trying to predict on to be effectively similar enough that we know things are in the same spot. The challenges come in whenever your training data is different from what you're trying to predict. So that could mean if all your training data is done in low light, and you have a really sunny day, then it's going to be really difficult to predict a match. If it's something as simple as you have on a pair of sunglasses, that removes a lot of the landmarks and makes it a lot harder to predict.

Well, the next step is to try to extract information from that faceand. We do that by passing the photo through a neural net each layer until the last one is going to try to extract features and then represent them in a mathematical form, so liz. Where are these programs getting the data to match these faces? Well, usually, a facial recognition system has a database of photos that have already been run through a neural net, meaning they're already translated into math into multi-dimensional coordinates. Now, if i have a database - and i know that this mathematical representation looks like liz o'sullivan, then i can also take this new photo and try to figure out how close that vector space is to the resulting vector space of my newphoto. If it's very close, then yes, it might be liz if they're very far apart. It'S likely it's somebody else. People are starting to realize exactly how powerful this tool is, and, in some cases, we're finding out that law enforcement and federal agencies have been using it. For some time, without our knowledge or consent, i'm not a lot of cases we're finding out that this technology doesn't work as well on people of color or women, or you know gender, fluid or non-binary people. So this is important now for us to take a moment to understand fully exactly what we want our civil liberties to look like in the us. Hey nbc news viewers thanks forchecking out our youtube channel subscribe by clicking on that button down here and click on any of the videos over here to watch. The latest interviews show highlights and digital exclusives thanks for watching.

Watch Next