Neighborhood Contrastive Learning for Novel Class Discovery

By Zhun Zhong, Enrico Fini, Subhankar Roy, Zhiming Luo, Elisa Ricci, Nicu Sebe. 2021

202 citations recorded by the source corpus.

In this paper, we address Novel Class Discovery (NCD), the task of unveiling new classes in a set of unlabeled samples given a labeled dataset with known classes. We exploit the peculiarities of NCD to build a new framework, named Neighborhood Contrastive Learning (NCL), to learn discriminative representations that are important to clustering performance. Our contribution is twofold. First, we find that a feature extractor trained on the labeled set generates representations in which a generic query sample and its neighbors are likely to share the same class. We exploit this observation to retrieve and aggregate pseudo-positive pairs with contrastive learning, thus encouraging the model to learn more discriminative representations. Second, we notice that most of the instances are easily discriminated by the network, contributing less to the contrastive loss. To overcome this issue, we propose to generate hard negatives by mixing labeled and unlabeled samples in the feature space. We experimentally demonstrate that these two ingredients significantly contribute to clustering performance and lead our model to outperform state-of-the-art methods by a large margin (e.g., clustering accuracy +13% on CIFAR-100 and +8% on ImageNet).

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

Citations in the corpus

Novel Class Discovery without Forgetting (2022, methodology): “– When compared with prominent class-discovery methods [16,19,60] adapted to our proposed setting, our methodology achieves improved class-discovery performance with significantly less forgetting.”
Class-incremental Novel Class Discovery (2022, methodology): “We split the data sets into the old and new classes following the existing NCD and iNCD works [15,36,29].”
Generalized Category Discovery (2022, methodology): “Meanwhile, self-supervised contrastive learning has been widely used as pre-training to achieve robust representations in NCD [21,49].”
Joint Representation Learning and Novel Category Discovery on Single- and Multi-modal Data (2021, background): “performance by augmenting the training samples by mixing the labelled and unlabelled data [52], aggregating pseudopositive pairs with contrastive learning [51], designing a unified objective [12], and leveraging local part information [50].”
Residual Tuning: Toward Novel Category Discovery Without Labels (2022, unknown): “Because of the performance gains, recent research works [17], [18] also introduce self-supervised pre-training in their methods.”
Novel Class Discovery in Semantic Segmentation (2021, background): “In NCDSS, we have a labeled base set [9, 40], the number of novel classes N uC is a known prior.”
Spacing Loss for Discovering Novel Categories (2022, methodology): “RS [11, 12], NCL [45] and OpenMix [46] first use RotNet [22] to self-supervise on the labeled and unlabeled data.”
Novel Visual Category Discovery with Dual Ranking Statistics and Mutual Knowledge Distillation (2021, unknown): “Concurrent to our work, several methods [28, 58, 57, 14] are proposed to improve the performance of novel category discovery from different perspectives, showing promising results.”
A Unified Objective for Novel Class Discovery (2021, methodology): “In particular, UNO still achieves state-of-the-art results on CIFAR10 while it is slightly outperformed by NCL [30] on CIFAR100-20 and ImageNet.”