How to identify faces using machine learning algorithms?

Question

Classification

For the purpose of self-education, I decided to start studying issues related to machine learning algorithms. To make the study interesting, I decided to try to implement the task of identifying faces in photographs. Some articles helped (for example, https://habrahabr.ru/post/317798/) and I learned how to highlight faces in photographs, align them, but there are problems with the last stage - classification. If you act on the bulk of the examples, then everything turns out well, since the finite number of classes and the model is trained and produces acceptable results. This is backstory.
My main question is how to make a classifier that can be trained to identify a new person. For example, there is a classifier for 10 outputs, how to create a new classifier that will be able to identify a new person, i.e. will have 10 + 1 output, but at the same time not completely retraining the classifier, but expanding the dimension and retraining to identify a new face?
Related question, is there another way to quickly map a feature vector to a person ID? Does Facebook have a classifier with 1.6 billion outputs that matches a face from a photo with an account?

Answer 1

You can train any neural network, in the case of images it makes sense: google transfer learning and fine tuning (for example, cs231n.github.io/transfer-learning/).

Answer 2

Instead of identifiers, try to get a fixed-dimensional vector as an output. Then, it remains for you to vectorize the faces already existing in the database, and compare the new ones that the network will issue. For example, through the consine distance.