Modeling Similarity: theory and application of Siamese networks

This thesis contributes to the research on Siamese networks by presenting how they can be used to devise a soft estimation of the Graph–Edit Distance, which allows to solve the NP–complete graph isomorphism problem in an affordable time. Then, three distinct applications of Siamese networks are presented, related to two prediction and one classification tasks. In all of these applications, the embedding space is constructed based on the similarity target, without relying on the classical Siamese loss functions, such as contrastive or triplet losses. These examples showcase how Siamese networks can build powerful embedding spaces even without traditional loss functions, underlining their flexibility and effectiveness in solving real–world tasks across different fields.