Generalized Category Discovery With Unknown Sample Generation.

Area of Science:

• Artificial Intelligence
• Machine Learning

Background:

• Semi-supervised learning (SSL) typically assumes test data belongs only to known classes.
• Real-world data often contains novel, previously unseen categories.
• Generalized Category Discovery (GCD) extends SSL to handle both known and unknown classes in unlabeled data.

Purpose of the Study:

• To propose a novel approach for Generalized Category Discovery (GCD) by generating unknown samples.
• To address the challenge of unknown category uncertainty in GCD.
• To enhance model robustness when encountering new classes.

Main Methods:

• Developed Generalized Category Discovery with Unknown Sample Generation (GCDUSG).
• Employed a prototype alignment method to estimate unknown category numbers and assign pseudo-labels.
• Generated realistic unknown samples by leveraging known-unknown prototype relationships and minimizing Maximum Mean Discrepancy.
• Incorporated a pseudo-label supervision loss for comprehensive classifier training.

Main Results:

• Demonstrated the effectiveness of the proposed GCDUSG approach.
• Achieved improved performance in handling datasets with both known and unknown classes.
• Validated the method's ability to generate discriminative unknown samples.

Conclusions:

• The proposed GCDUSG method offers a viable solution for Generalized Category Discovery.
• Generating unknown samples is a promising strategy for improving model adaptability to novel classes.
• The approach enhances classifier performance in real-world scenarios with evolving data distributions.