How can artificial intelligence recognize a face? This is how.
Facial recognition on a large scale became possible thanks to the breakthrough of deep machine learning at the beginning of this decade.
Deep machine learning or deep neural networks are about a computer program that learns on its own. The fact that it is called “neural” or “neural network” comes from the fact that the technology is inspired by the human brains’ properties to transform data into information. It is a variant of the broader concept of machine learning, which in turn is part of what we call artificial intelligence.
With deep machine learning, an algorithm serves training data and spits out a result. But on the way between these two, the algorithm interprets the signals – thus training data – in a number of layers. For each new layer, the degree of abstraction increases
Say that you want to build a deep neural network that can differentiate different faces or that can determine which faces are identical. Exercise data should then be a large number of images on the faces (the larger the dataset, the more accurate the network, at least in theory). Partly on different faces, but also many pictures on the same face.
Measurements are unique to each face
Of course, a computer does not “see” a face in the image, but a number of values representing different pixels. With the pixels as a background, the deep neural network learns to find patterns. For each layer passed in the network, some patterns become more interesting (stronger signal between the “neurons” in the network) while others are ignored (weaker signal). During training, what AI researchers call the “weights” of the various signals to produce the desired result better and better.
The first, second and hundredth time the algorithm performs this procedure, the results are usually not as good, but eventually, the network can achieve impressive results. In a way, it can be said that the network has learned to abstract and generalize, from raw pixel values to the classification of different people’s faces.
But that is perhaps not what we humans think of when we use terms such as generalization: it is rather that the network has worked out a number of metrics that are unique to each face. If the pre-trained network is served a new image on a face, the network can match that face’s measurement values to the faces on other images. If the network generates roughly the same values for different images, it is likely the same person on both images.
It is called deep machine learning because such a model can use multiple – sometimes hundreds – layers. There is also a symbolic meaning in such a way that we humans cannot really understand how the computer program does to find patterns. It operates, so to speak, deep, beneath the surface.
Although the algorithms are evolving and refining as they are, there are really two other reasons behind the breakthrough of deep neural networks: access to large data sets and cheap computing power, especially in the form of graphics cards that were most often associated with computer games.
It may also be borne in mind that the method described above for classification purposes is only one of many. But it is common.