Deep Meta Metric Learning via Learnable Distance

In this thesis work, we propose a Deep Metric Learning method via learnable distance to solve image retrieval problems. Unlike the conventional approaches where a certain predefined distance is used (such as the most commonly Euclidean distance) to compute the dissimilarity between embeddings, what we propose is to learn this distance with a neural network. This learned distance has the objective of performing better than a certain target distance (such as the Euclidean distance), trying to put the examples belonging to the same class closer and the examples belonging to different classes farther away in the embeddings space. The training of the network that produces the distance occurs simultaneously with the training of the embeddings space following the Meta Learning literature. In particular, in the training algorithm we update the distance with a “inner optimization” of the distance network. Then this updated distance is given as input to the “outer optimization” of the network that produces the embeddings. This work also wants to prove the quality of the proposed solution by comparing the results obtained with the state of the art. Experiments results on the widely used CUB-200-2011 and Stanford Online Products datasets want to demonstrate the effectiveness of the proposed approach.