Stochastic Gradient Descent for Kernel-Based Maximum Correntropy Criterion
Author Information
Author(s): Li Tiankai, Wang Baobin, Peng Chaoquan, Yin Hong, Trapp Martin, Alquier Pierre
Primary Institution: School of Mathematics and Statistics, South-Central MinZu University, Wuhan, China
Hypothesis
Can stochastic gradient descent (SGD) be effectively applied to the kernel version of the maximum correntropy criterion (MCC) for robust learning?
Conclusion
The study establishes a theoretical foundation for applying SGD to MCC, demonstrating that it can achieve global linear convergence rates.
Supporting Evidence
- The study provides a rigorous analysis of the convergence behaviors of SGD applied to kernel MCC.
- It shows that the choice of step size can lead to global linear convergence rates.
- The results refine previous work and establish a theoretical foundation for SGD with MCC.
Takeaway
This study shows how a special math method called SGD can help make predictions better, even when the data is messy.
Methodology
The paper analyzes the convergence properties of SGD applied to kernel-based MCC and provides explicit convergence rates under certain conditions.
Limitations
The optimal selection of the robustness parameter σ in practice remains unknown.
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website