Deep Learning and the Future of Face Recognition

The success of AlexNet in large scale image recognition suggests that we may be entering a very different era of computer vision. Instead of carefully designing every feature by hand, deep neural networks can begin to learn useful visual representations directly from large collections of images. If this approach can be adapted to faces, it may allow recognition systems to become far more robust to changes in lighting, pose, expression and even partial occlusion.

The implications could be significant. A sufficiently powerful facial recognition system might authenticate a person from an ordinary camera image and could eventually be used for secure access, identity verification and border control. With larger datasets and increasing computing power, it may become possible to compare faces at a scale and speed that conventional methods cannot easily achieve.

But greater capability also brings greater responsibility. Facial recognition deals directly with identity, so questions of privacy, consent, accuracy and inappropriate surveillance cannot be treated as secondary concerns. If deep learning is to make face recognition dramatically more powerful, the technical challenge will be only part of the problem. We will also need to decide where such systems should be used and what safeguards must accompany them.