Convolutional Prototype Network for Open Set Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional Convolutional Neural Networks (CNNs) excel in closed-set recognition (CSR) but struggle with unknown data in open environments.
• Lack of robustness in CNNs for open-set recognition (OSR) necessitates new frameworks for real-world applications.

Purpose of the Study:

• To develop a deep learning framework that improves CNN robustness in open-set recognition (OSR) while preserving accuracy in closed-set recognition (CSR).
• To introduce a novel Convolutional Prototype Network (CPN) that integrates human-like prototype learning with CNN representation.

Main Methods:

• Proposed the Convolutional Prototype Network (CPN), replacing the traditional softmax with a prototype model for open-world scenarios.
• Designed discriminative losses to enhance CPN's classification of known samples.
• Introduced a generative loss, inspired by generative models, to improve robustness against unknown data and act as latent regularization.

Main Results:

• CPN demonstrated effectiveness as a hybrid model, combining advantages for both CSR and OSR tasks.
• End-to-end training jointly optimized the convolutional network and prototypes.
• Two novel rejection rules were proposed for detecting different types of unknowns in OSR applications.

Conclusions:

• The Convolutional Prototype Network (CPN) offers a robust and effective solution for open-set recognition challenges.
• Experimental results validate the efficiency and effectiveness of CPN across various datasets for both CSR and OSR.