This paper introduces a deep metric learning algorithm that leverages structured feature embeddings to enhance the performance of neural networks in learning semantic feature embeddings. The authors propose a method to lift the pairwise distances within a training batch to a matrix of pairwise distances, enabling the optimization of a novel structured prediction objective. This approach significantly improves the embedding quality compared to existing methods, as demonstrated on the CUB-200-2011, CARS196, and Online Products datasets. The Online Products dataset, with 120k images and 23k classes, is one of the largest publicly available datasets for metric learning. The proposed method not only enhances clustering and retrieval quality but also simplifies the network structure, making it more efficient and robust. The paper also discusses related work in deep metric learning, deep feature embedding, and zero-shot learning, and provides detailed implementation and evaluation results.This paper introduces a deep metric learning algorithm that leverages structured feature embeddings to enhance the performance of neural networks in learning semantic feature embeddings. The authors propose a method to lift the pairwise distances within a training batch to a matrix of pairwise distances, enabling the optimization of a novel structured prediction objective. This approach significantly improves the embedding quality compared to existing methods, as demonstrated on the CUB-200-2011, CARS196, and Online Products datasets. The Online Products dataset, with 120k images and 23k classes, is one of the largest publicly available datasets for metric learning. The proposed method not only enhances clustering and retrieval quality but also simplifies the network structure, making it more efficient and robust. The paper also discusses related work in deep metric learning, deep feature embedding, and zero-shot learning, and provides detailed implementation and evaluation results.