Deep neural networks give excellent results in many visual object recognition and image classification tasks. However, to achieve good performance, the network must be trained using very large number of examples from each category. In many practical situations, this is not always feasible. For example, for face recognition applications, we have only few or even one training example per each person.

In this talk, I’ll present a class of neural network architectures, such as Siamese or triplet networks, based on ‘metric learning’ or ‘similarity learning’ concept. These methods can be effectively used to learn from limited number of examples from each category. I’ll demonstrate how these methods can be used to solve various practical problems, such as face recognition, pedestrian re-identification or visual search.