Proxy Based Deep Metric Learning

Proxy-based deep metric learning aims to learn effective distance metrics in high-dimensional spaces by using "proxies" – representative embeddings for each class – to guide the learning process. Current research focuses on improving proxy-based methods by addressing issues like aligning sample and proxy distributions, incorporating uncertainty modeling (e.g., using probabilistic distributions), and enhancing the representation of intra-class structure through techniques such as non-isotropy regularization. These advancements lead to improved performance in various applications, including image retrieval, entity linking, and few-shot learning across domains like computer vision and natural language processing.