Abstract:Few-shot open-set recognition (FSOR) is a challenging task that requires a model to recognize known classes and identify unknown classes with limited labeled data. Existing approaches, particularly Negative-Prototype-Based methods, generate negative prototypes based solely on known class data. However, as the unknown space is infinite while the known space is limited, these methods suffer from limited representation capability. To address this limitation, we propose a novel approach, termed \textbf{D}iversified \textbf{N}egative \textbf{P}rototypes \textbf{G}enerator (DNPG), which adopts the principle of "learning unknowns from unknowns." Our method leverages the unknown space information learned from base classes to generate more representative negative prototypes for novel classes. During the pre-training phase, we learn the unknown space representation of the base classes. This representation, along with inter-class relationships, is then utilized in the meta-learning process to construct negative prototypes for novel classes. To prevent prototype collapse and ensure adaptability to varying data compositions, we introduce the Swap Alignment (SA) module. Our DNPG model, by learning from the unknown space, generates negative prototypes that cover a broader unknown space, thereby achieving state-of-the-art performance on three standard FSOR datasets.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the problem of insufficient representation of unknown - class samples in the **Few - Shot Open - Set Recognition (FSOR) task**. Specifically, the existing Negative Prototypes (NPs) generation methods mainly rely on known - class data to generate negative prototypes, which leads to a lack of diversity in negative prototypes and an inability to fully represent samples in the unknown space. ### Problem Background The FSOR task requires the model to be able to recognize known - class samples and correctly identify unknown - class samples with only a small amount of labeled data. Existing methods, especially those based on negative prototypes, usually generate negative prototypes using only known - class data. However, since the unknown space is infinite and the known space is finite, the negative prototypes generated by these methods are often limited to the characteristics of the known classes and cannot effectively represent the diversity of the unknown classes. For example, for the "Dalmatian" class, the existing negative prototypes may generate samples that are similar to dogs but with slightly different features, and cannot represent a broader unknown space, such as a guitar or other objects unrelated to Dalmatians. ### Solution Proposed in the Paper To solve this problem, the paper proposes a new method named **Diversified Negative Prototypes Generator (DNPG)**. The core idea of DNPG is to learn unknown information from the unknown space to generate more diverse negative prototypes. Specifically: 1. **Learn inverse representations from known classes**: In the pre - training stage, by learning the inverse representation (called open weights) of each base class, capture the features that do not belong to this class, thereby representing the corresponding unknown space. 2. **Generate negative prototypes using open weights**: In the meta - learning stage, use these open weights and the similarity relationships between base classes and new classes to generate more diverse negative prototypes. 3. **Introduce the Swap Alignment module**: In order to prevent the collapse of negative prototypes and ensure that they adapt to different data distributions, the Swap Alignment module is introduced to ensure that the generated negative prototypes can distinguish different unknown - class samples. ### Summary The main contributions of the paper include: - Proposing a new FSOR method - DNPG, which generates diverse negative prototypes by learning information in the unknown space. - Introducing the Swap Alignment module to prevent the collapse of negative prototypes and enhance the adaptability to new - class data. - Conducting experiments on multiple widely - used FSOR datasets, proving that the DNPG model is superior to the current state - of - the - art methods. Through these improvements, DNPG can more effectively identify unknown - class samples under the condition of few - shot, and improves the performance of the FSOR task.

Learning Unknowns from Unknowns: Diversified Negative Prototypes Generator for Few-Shot Open-Set Recognition

Convolutional Prototype Network for Open Set Recognition

Overall positive prototype for few-shot open-set recognition

Zero-Shot Learning with Generative Latent Prototype Model.

Learning Relative Feature Displacement for Few-Shot Open-Set Recognition

Learning Multiple Gaussian Prototypes for Open-Set Recognition

Exploring Diverse Representations for Open Set Recognition

Glocal Energy-based Learning for Few-Shot Open-Set Recognition

Few-shot Open Relation Extraction with Gaussian Prototype and Adaptive Margin

Understanding Negative Proposals in Generic Few-Shot Object Detection

P-ODN: Prototype based Open Deep Network for Open Set Recognition

Open-Set Recognition of Screen Defects with Negative-Guided Augmented Prototype Generator and Open Feature Generation

Learning adversarial semantic embeddings for zero-shot recognition in open worlds

Generative Adversarial Network and Mutual-Point Learning Algorithm for Few-Shot Open-Set Classification of Hyperspectral Images

Adversarial Reciprocal Points Learning for Open Set Recognition

Synergetic proto-pull and reciprocal points for open set recognition

Few-shot Open-set Recognition by Transformation Consistency

Know Yourself Better: Diverse Discriminative Feature Learning Improves Open Set Recognition

Open-set long-tailed recognition via orthogonal prototype learning and false rejection correction

Open Long-Tailed Recognition In A Dynamic World