CSRP: Modeling class spatial relation with prototype network for novel class discovery

Abstract

Novel Class Discovery(NCD) is a learning paradigm within the open-world task, in which machine learning models leverage prior knowledge to guide unknown samples into semantic clusters in an unsupervised environment. Recent research notes that maintaining class relations can assist classifiers in better recognizing unknown classes. Inspired by this study, we propose Class-Spatial-Relation modeling with a Prototype network (CSRP). A prototype network is a machine learning model used to classify tasks. It performs by learning prototypes for each class and makes classification decisions based on the similarity between a given sample and these prototypes. It conducts complex class boundaries better than linear classification models, providing higher flexibility and accuracy for classification tasks. Specifically, the proposed prototype network enables spatial modeling based on the distance between samples and each prototype, which can better obtain class relation information to improve the model’s interpretability and robustness. In addition, we simultaneously perform knowledge distillation on known and unknown classes to balance the model’s classification performance for each class. To evaluate the effectiveness and generality of our method, we perform extensive experiments on the CIFAR-100 dataset and fine-grained datasets: Stanford Cars, CUB-200-2011, and FGVC-Aircraft, respectively. Our method results are comparable to existing state-of-the-art performance in the standard dataset CIFAF100, while outstanding performance on three fine-grained datasets surpassed the baseline by 3%-9%. In addition, our method creates more compact clusters in the latent space than in linear classification. The success demonstrates the effectiveness of our approach.