KMT-PLL: K-Means Cross-Attention Transformer for Partial Label Learning

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Partial Label Learning (PLL) addresses instance classification with multiple candidate labels, where only one is correct.
• Vision Transformers (ViT) show promise in clean data but face challenges in PLL applications.
• Existing methods struggle with the inherent noise and ambiguity in partial label datasets.

Purpose of the Study:

• To develop a novel approach for improving Vision Transformer performance in Partial Label Learning.
• To effectively handle noisy labels and disambiguate instances within partial label settings.
• To introduce a method for continuous learning of cluster centers representing label classes.

Main Methods:

• Proposed K-means cross-attention transformer for PLL (KMT-PLL) to learn cluster centers.
• Utilized K-means cross-attention for effective representation of label classes and noise detection.
• Introduced a corrected cross-entropy formulation weighting candidate labels based on instance-to-label relevance.

Main Results:

• KMT-PLL effectively learns cluster centers, enabling measurable similarity between instances and labels.
• The method successfully detects and mitigates the impact of noisy labels during training.
• Simulation results validate the significant advantages and suitability of KMT-PLL for PLL tasks.

Conclusions:

• KMT-PLL offers a robust solution for Partial Label Learning by refining instance-to-label relevance.
• The proposed approach enhances classifier training through progressively identified ground-truth labels and refined cluster centers.
• KMT-PLL demonstrates superior performance and adaptability in challenging partial label learning scenarios.