Interpretable inverse iteration mean shift networks for clustering tasks.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Deep Learning

Background:

• Neural networks excel at feature representation but lack interpretability.
• The Mean Shift algorithm offers interpretability but has limited representation power.

Purpose of the Study:

• To develop a novel architecture, the Mean Shift Network (MS-Net), that combines the strengths of neural networks and the Mean Shift algorithm.
• To enhance interpretability in deep learning models while maintaining strong feature representation capabilities.

Main Methods:

• Proposed MS-Net, an inverse iteration fuzzy clustering network where each layer is interpretable.
• Introduced a continuously-differentiable Gaussian-inspired kernel for the membership layer to ensure convergence.
• Developed a weighted version (WMS-Net) to account for training example importance.
• Considered curvature-based extensions (CB-MS-Net, CB-WMS-Net).

Main Results:

• MS-Net and its variants demonstrate strong feature representations with inherent interpretability.
• Theoretical results with proofs of weak and strong convergence were presented.
• Simulation results on 11 datasets (5 clustering, 6 real-world) confirmed the effectiveness of the proposed algorithms.

Conclusions:

• MS-Net offers a promising approach to bridge the gap between interpretability and representation power in deep learning.
• The proposed architecture and its extensions are effective for clustering tasks and real-world applications.