Recently, the 34th Conference on Neural Information Processing Systems (NeurIPS) announced the results of paper acceptance. The paper Deep Metric Learning with Spherical Embedding under the guidance of Associate Professor Li Yingming, with a doctoral student Zhang Dingyi as the first author, was accepted.

NeurIPS of this year received a total of 9,454 valid submissions and 1,900 papers, with an acceptance rate of about 20.1%, hitting a record low. NeurIPS is an annual interdisciplinary conference that mainly includes artificial intelligence and natural neural information processing. It is not only an internationally recognized top conference in the field of machine learning and artificial intelligence research but also a class A conference of the Chinese Computer Society (CCF) receiving widespread attention from academia and industry

The paper studies the optimization of features in angular space in metric learning, analyzes the specific impact of feature normalization in model optimization for loss functions based on sample pairs, and reveals the negative effects of inconsistent feature norms on model updates. Based on this, a spherical feature constraint (SEC) is proposed to adaptively adjust the feature norm and improve the model optimization process, as shown in Figure 1.

Figure 1Adjustment of angle update

This method is simple to implement and easily combined with existing loss functions, which can improve the effect of existing methods. The results on metric learning, face recognition, and comparative self-supervised learning tasks show that SEC not only improves performance of the existing model but also significantly accelerates the convergence speed of the model.