Parameter-Free Deep Multi-Modal Clustering With Reliable Contrastive Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Deep multi-modal clustering (DMC) leverages diverse data sources for improved performance.
• Heterogeneous data distributions and quality variations across modalities pose challenges for existing DMC methods.
• Contrastive learning in DMC can be hindered by learning from low-quality or unreliable modalities.

Purpose of the Study:

• To propose a novel parameter-free deep multi-modal clustering framework (PDMC-RCL).
• To address the limitations of uneven data quality in multi-modal clustering.
• To enhance feature representation learning through reliable, multi-level contrastive learning.

Main Methods:

• Introduced parameter-free deep multi-modal clustering with reliable contrastive learning (PDMC-RCL).
• Developed a reliable contrastive learning mechanism to quantify modality pair relationships using weights.
• Implemented multi-level contrastive learning at both feature and cluster levels.

Main Results:

• PDMC-RCL effectively handles heterogeneous data distributions and uneven quality.
• The reliable contrastive learning selectively promotes learning from useful modality pairs.
• Experimental results demonstrate superior performance compared to state-of-the-art DMC methods across various datasets.

Conclusions:

• PDMC-RCL offers a robust and parameter-free solution for multi-modal clustering.
• The proposed reliable contrastive learning strategy significantly improves feature representation and clustering accuracy.
• The method achieves promising results without requiring additional hyperparameter tuning.