Facial recognition systems are image classifiers where the classes are persons, represented as sets of images of their faces. Each person is assigned a numerical id as a class label and classification means that the system matches an image to a label.
Such systems are used in one of two modes, verification or identification.
Verificatiom means that the system is given as input an image and a class label and outputs positive if the input matches the label, and negative otherwise.
Identification means that the system is given as input an image and outputs a class label.
In either case, the system may not directly return a single label, but a set of labels each associated to a real-valued number, interpreted (by the operators of the system) as a likelihood. However, in that case the system has a threshold delimiting positive from negative identifications. That is, if the likelihood that the system assigns to a classification is above the threshold, that is considered a "positive identificetion", etc.
In other words, yes, a system that outputs a continuous distribution over classes representing sets of images of peoples' faces can still be "wrong".
Think about it this way: if a system could only ever signal uncertainty, how could we use it to make decisions?
Such systems are used in one of two modes, verification or identification.
Verificatiom means that the system is given as input an image and a class label and outputs positive if the input matches the label, and negative otherwise.
Identification means that the system is given as input an image and outputs a class label.
In either case, the system may not directly return a single label, but a set of labels each associated to a real-valued number, interpreted (by the operators of the system) as a likelihood. However, in that case the system has a threshold delimiting positive from negative identifications. That is, if the likelihood that the system assigns to a classification is above the threshold, that is considered a "positive identificetion", etc.
In other words, yes, a system that outputs a continuous distribution over classes representing sets of images of peoples' faces can still be "wrong".
Think about it this way: if a system could only ever signal uncertainty, how could we use it to make decisions?