Few-shot Novel Category Discovery

TL;DR

FSNCD addresses the gap between few-shot learning and open-set novel category discovery by defining episodes with $N$-way $K$-shot support where the query set may contain novel categories beyond the labeled base classes; a two-phase training pipeline uses supervised contrastive learning for robust representations and introduces SHC and UKC to flexibly cluster unknown categories without fixed counts. Experiments across five datasets with a ViT-DINO backbone show that SHC and UKC deliver leading performance on both old and new classes, with UKC providing strong results on large-scale data and real-time inference trade-offs. This work advances open-set recognition under label-scarce, realistic conditions and provides scalable clustering strategies for dynamic category discovery in vision tasks.

Abstract

The recently proposed Novel Category Discovery (NCD) adapt paradigm of transductive learning hinders its application in more real-world scenarios. In fact, few labeled data in part of new categories can well alleviate this burden, which coincides with the ease that people can label few of new category data. Therefore, this paper presents a new setting in which a trained agent is able to flexibly switch between the tasks of identifying examples of known (labelled) classes and clustering novel (completely unlabeled) classes as the number of query examples increases by leveraging knowledge learned from only a few (handful) support examples. Drawing inspiration from the discovery of novel categories using prior-based clustering algorithms, we introduce a novel framework that further relaxes its assumptions to the real-world open set level by unifying the concept of model adaptability in few-shot learning. We refer to this setting as Few-Shot Novel Category Discovery (FSNCD) and propose Semi-supervised Hierarchical Clustering (SHC) and Uncertainty-aware K-means Clustering (UKC) to examine the model's reasoning capabilities. Extensive experiments and detailed analysis on five commonly used datasets demonstrate that our methods can achieve leading performance levels across different task settings and scenarios.

Figures (6)

• Figure 1: Comparison of Different Task Settings. Solid lines represent labeled data, dashed lines represent unlabeled data, with different colors indicating different categories. Unlabeled data in test phase are samples to be classified. Inductive learning and transductive learning are annotated as "Train/Test" and "Train+Test". Please note that during the generalization testing phase in inductive learning, FSL and FSNCD are based on a small amount of labeled data.
• Figure 2: Illustration of our proposed baselines. In the representation learning stage, each episode takes both support and query samples into a min-batch selected from $\mathcal{D}^\text{L}$ and trains a feature extractor by using the supervised contrastive learning approach. In the second stage, we test on $\mathcal{D}^{U}$, and for each episode, we utilize the proposed clustering method to classify samples belonging to $\mathcal{Y}^\text{S}$ and cluster new categories.
• Figure 3: Ablation studies for quantity of query using the 5w5s configuration. (a) The accuracy for Imagenet-100 on All classes. (b) The accuracy for CUB-200 All classes.
• Figure 4: An Example of Uncertainty-aware K-means Clustering (UKC) Based FSNCD. In the initialization stage, the model is clustered into $\left|\mathcal{Y_S}\right|$ classes with random centroids. The proposed criteria are employed to split clusters, extracting category centroids as heuristic information for the next iteration. This process continues until the specified conditions are met, resulting in the predictions of model.
• Figure 5: Dilemmas faced by real-time reasoning tasks. The pentagon represents prototypes supporting samples, the solid line rhombus represents a query sample, and the dashed line rhombus represents a potential distribution that may exist.