The siamese network is a special architecture originally used for signature verification and one-shot classification. To perform this task, it predicts whether two images are in the same class.
As indicated in the diagram above, the two images go through the same network with identical architecture and weights. Their highest-level representations can then be processed in two ways:
- Concatenate them and use another neural module to predict probability of the same class. This can be trained with 💧 Cross Entropy loss.
- Directly compute their similarity via some distance metric and train with one of the Contrastive Objectives.