Facebook’s DeepFace Software Is Almost As Good As Humans at Matching Faces
Do you ever wonder what Facebook does with its trove of party pictures and Instagram shots? Well, besides storing photo albums that will only serve give our kids and grandkids a good laugh, Facebook has assembled an artificial intelligence team that’s using our photos to teach computers to verify human faces. According to a recent paper released by Facebook, the team has developed software that can match faces in photographs almost as well as humans can.
Using Deep Learning to Match Faces
The project, called DeepFace, employs an area of artificial intelligence called deep learning to match pictures of the same person among photos of millions of faces, no matter what angle the photos are taken from or how varied the lighting is. In a test using 4 million photos of faces belonging to almost 4,000 Facebook users, the software correctly matched photos of the same face 97.25% of the time, while human beings did so 97.52% of the time. This is a significant improvement over existing face-matching software which mistook faces 25% more often than DeepFace.
DeepFace software works to match two photographs that show the same face. This is different from the facial recognition tool Facebook currently uses which works to put a name to a face. It’s possible, though, that some of the techniques behind the DeepFace software could eventually make Facebook’s current facial recognition software more accurate, according to MIT Technology Review.
How DeepFace Works
To process facial images, DeepFace first corrects the angle of a face so that the person is facing forward. Next a simulated neural network converts the image into a numerical description, and, if it is close enough to the numerical description of another image, DeepFace determines the two images show the same face.
In order to be applied more widely, DeepFace would have to be trained on an even more extensive album of faces (and Facebook has no shortage of albums). Afterward the software could potentially be used to automate photo tagging on Facebook, or at least to help Facebook better understand the ways in which Facebook friends interact with each other (say hello to more targeted advertising?). For now, though, the software hasn’t found an application outside the research setting.