The high-level solution to the problems is through the similarity function, which states the following:

• d(face_1, face_2): Degree of difference between images
• d(face_1, face_2) <: Means same face else, not same

This means, instead of trying to learn to recognize a specific person's face, what if we learn a function, d, which measures how similar or different two images are. If the function were to return a value smaller than a constant, such as gamma, then we say this is the same face or the same image. Otherwise, we say, they are different.

Suppose on the left we have the employee's faces, and on the right there is a person coming in:

We will have numbers returned by the d function, and in the first two cases, we see that we have big numbers, because actually these two images, and these two people are different. Well, the third case had a really small number, and if we have a constant such as 0.8 or something, then, we have the function output that these images are the same. And for the fourth case, naturally we have a high number.

Additionally, this solution also scales well, since if a new person is joining, that would mean just a new comparison to execute. We don't have any new neural network to train or such thing because the d function is quite general. We are learning a function to measure how similar two images are, rather than learning a specific face.

Let's now move on to searching solutions for face recognition.