ReCL: A Plug-and-Play Module for Enhancing Generalized Category Discovery Using Transport-Based Method to Uncover the...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Deep learning models struggle in open-set recognition due to differing training and testing label sets.
• Generalized Category Discovery (GCD) addresses this by classifying both labeled and unlabeled data simultaneously.
• Current contrastive learning methods for GCD often neglect inter-sample relationships within unlabeled data.

Purpose of the Study:

• To propose a novel relationship-based contrastive learning (ReCL) module to enhance Generalized Category Discovery.
• To improve the learning of discriminative features for unlabeled samples in GCD tasks.
• To integrate ReCL with existing GCD models to boost their performance.

Main Methods:

• Developed a relationship-based contrastive learning (ReCL) module for Generalized Category Discovery.
• Employed a transport-based assignment method to identify semantically relevant samples for unlabeled data points.
• Utilized a prototype-based fusion technique to create a positive anchor for contrastive learning with unlabeled data.

Main Results:

• The proposed ReCL module significantly improves the performance of various existing GCD models.
• Experiments across diverse domains show state-of-the-art results on multiple benchmarks.
• The transport-based sample selection in ReCL yields semantically similar yet diverse samples compared to cosine similarity methods.

Conclusions:

• The relationship-based contrastive learning module effectively addresses limitations in current GCD approaches.
• ReCL offers a flexible and powerful addition to existing GCD frameworks, enhancing their capabilities.
• The method demonstrates superior performance by capturing nuanced relationships within unlabeled data.