Date

2022-8-24

Author

Gürbüz, Yeti Ziya

Metadata

Show full item record

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Item Usage Stats

247
views

209
downloads

Cite This

Deep metric learning (DML) aims to fit a parametric embedding function to data of semantic information (e.g. images) so that l2-distance between embedded samples is low whenever they share similar semantic entities. An embedding function of such behavior is attained by minimizing empirical expected pairwise loss that penalizes inter-/intra-class proximity violations in embedding space. Proxy-based methods which use a learnable embedding vector per class in their loss formulation are state-of-the-art. We first address characterizing generalization error of proxy-based methods. We reformulate DML as a chance-constrained optimization problem and through careful theoretical analysis, we show that DML with better generalization guarantees can be achieved by iteratively minimizing a proxy-based loss and re-initializing proxies with embeddings of new samples. Second, we consider critical desideratum for DML: generalization to unseen data. We analyze global average pooling (GAP) which is an effective architectural choice to aggregate information in DML. With theoretical and empirical supports, we explain effectiveness of GAP by considering each feature vector as representing a different semantic entity and GAP as a convex combination of them. Following this perspective, we generalize GAP and propose a learnable generalized sum pooling method (GSP) improving GAP with two distinct abilities: i) the ability to choose a subset of semantic entities, effectively learning to ignore nuisance information, and ii) learning the weights corresponding to the importance of each entity. We further propose a zero-shot loss to ease the learning of GSP. We show the effectiveness of our contributions with extensive evaluations on 4 popular DML benchmarks.

Subject Keywords

metric learning, alternating projections, feature selection, zero-shot loss

URI

https://hdl.handle.net/11511/98619

Collections

Graduate School of Natural and Applied Sciences, Thesis