Cross-Modal Clustering With Deep Correlated Information Bottleneck Method.

Area of Science:

• Machine Learning
• Data Science
• Artificial Intelligence

Background:

• Cross-modal clustering (CMC) aims to improve clustering accuracy by leveraging correlations between different data modalities.
• Existing methods struggle with high-dimensional, non-linear data and conflicting information across modalities.
• Modality-private information can interfere with correlation mining, hindering clustering performance.

Purpose of the Study:

• To develop a novel method, Deep Correlated Information Bottleneck (DCIB), for effective cross-modal clustering.
• To address challenges in capturing inter-modal correlations and eliminating modality-private information.
• To enhance clustering accuracy by focusing on shared representations.

Main Methods:

• DCIB employs a two-stage data compression approach to eliminate modality-private information.
• It preserves correlations by focusing on shared representations across multiple modalities.
• The method utilizes a mutual information-based objective function with a variational optimization approach.

Main Results:

• Experimental results on four cross-modal datasets demonstrate the superiority of the DCIB method.
• The DCIB method effectively captures correlations between multiple modalities.
• Modality-private information is successfully eliminated, leading to improved clustering accuracy.

Conclusions:

• DCIB offers a robust solution for cross-modal clustering by effectively managing inter-modal correlations and modality-specific noise.
• The proposed method achieves superior performance compared to existing approaches.
• The study provides a valuable contribution to the field of multi-modal data analysis.