Robust Learning With Kernel Mean -Power Error Loss.

Area of Science:

• Machine Learning
• Statistical Signal Processing
• Data Science

Background:

• Correntropy is a robust second-order statistical measure used in kernel space for learning and signal processing.
• Existing methods may not fully capture complex data distributions for optimal robustness.

Purpose of the Study:

• Introduce a novel non-second order statistical measure in kernel space: kernel mean-power error (KMPE).
• Develop robust learning algorithms by applying KMPE to established machine learning models.
• Evaluate the performance enhancement offered by KMPE-based algorithms.

Main Methods:

• Defined kernel mean-power error (KMPE) as a non-second order statistical measure.
• Integrated KMPE into extreme learning machine (ELM) to create ELM-KMPE.
• Applied KMPE to principal component analysis (PCA) to develop PCA-KMPE.
• Conducted experiments using synthetic and benchmark datasets.

Main Results:

• The proposed KMPE measure encompasses correntropic loss (C-Loss) as a specific instance.
• Developed ELM-KMPE and PCA-KMPE algorithms demonstrate improved performance.
• Experimental validation confirms the efficacy of KMPE-based methods over existing approaches.

Conclusions:

• Kernel mean-power error (KMPE) offers a powerful new statistical framework for robust machine learning.
• KMPE-based ELM and PCA algorithms provide superior performance in data analysis tasks.
• The developed methods show significant potential for advancing robust learning and signal processing.