Hyperspherical Classification with Dynamic Label-to-Prototype Assignment
This work addresses a specific bottleneck in metric learning for classification, offering an incremental improvement over static assignment methods.
The paper tackles the challenge of capturing inter-class relationships in non-parametric classifiers by optimizing label-to-prototype assignments during training, resulting in improved accuracy of 1.22% on CIFAR-100 and 2.15% on ImageNet-200 with a smaller metric space dimension.
Aiming to enhance the utilization of metric space by the parametric softmax classifier, recent studies suggest replacing it with a non-parametric alternative. Although a non-parametric classifier may provide better metric space utilization, it introduces the challenge of capturing inter-class relationships. A shared characteristic among prior non-parametric classifiers is the static assignment of labels to prototypes during the training, ie, each prototype consistently represents a class throughout the training course. Orthogonal to previous works, we present a simple yet effective method to optimize the category assigned to each prototype (label-to-prototype assignment) during the training. To this aim, we formalize the problem as a two-step optimization objective over network parameters and label-to-prototype assignment mapping. We solve this optimization using a sequential combination of gradient descent and Bipartide matching. We demonstrate the benefits of the proposed approach by conducting experiments on balanced and long-tail classification problems using different backbone network architectures. In particular, our method outperforms its competitors by 1.22\% accuracy on CIFAR-100, and 2.15\% on ImageNet-200 using a metric space dimension half of the size of its competitors. Code: https://github.com/msed-Ebrahimi/DL2PA_CVPR24