Positive-negative prototypes fusion framework for open set recognition

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Open Set Recognition (OSR) faces a key challenge in balancing classification accuracy for known data with the identification of unknown data.
• Existing methods struggle to effectively manage the risks associated with the 'open space' where unknown data resides.

Purpose of the Study:

• To develop novel frameworks that improve the balance between classifying known data and managing unknown data risks in OSR.
• To enhance the generalization performance of OSR models and reduce open space risk.

Main Methods:

• Introduced the Positive-Negative Prototypes Fusion Framework (PNPFF) using multiple positive and single negative prototypes.
• Developed adversarial (APNPFF) and enhanced (APNPFF++) versions utilizing Generative Adversarial Networks (GANs) and manifold mix-up techniques.
• These methods aim to improve intra-class compactness and inter-class separation while simulating unknown data characteristics.

Main Results:

• PNPFF reduces classification risk for known data and partially alleviates open space risk by reserving space for unknowns.
• APNPFF and APNPFF++ further reduce open space risk by simulating unknown data, significantly enhancing generalization performance.
• Experiments across multiple benchmark datasets confirm the effectiveness and robustness of the proposed methods.

Conclusions:

• The proposed PNPFF, APNPFF, and APNPFF++ frameworks offer a superior approach to Open Set Recognition.
• These methods demonstrate significant improvements in handling both known and unknown class recognition tasks.
• The study highlights the potential of prototype-based and generative approaches for advancing OSR.