Single Hidden Layer Neural Networks With Random Weights Based on Nondifferentiable Functions.

Area of Science:

• Machine Learning
• Computational Science
• Artificial Intelligence

Background:

• Nondifferentiable functions are crucial in machine learning.
• Random-weight neural networks, including RVFL and ELM, benefit from efficient objective functions.

Purpose of the Study:

• To develop a novel framework for integrating nondifferentiable functions into RVFL and ELM objective functions.
• To enhance computational efficiency and maintain accuracy in machine learning models.

Main Methods:

• Incorporated six nondifferentiable functions (norms $L_{1,1}$, $L_{1,2}$, $L_{2,2}$, AbsMin, AbsMax, MaxMin) into RVFL and ELM objective functions.
• Utilized Fourier random assignments as activation functions for enhanced robustness.
• Evaluated algorithms on 12 benchmark datasets against $L_{2,1}$-RF-ELM.

Main Results:

• Algorithms utilizing nondifferentiable functions achieved high accuracy across various dataset sizes.
• Demonstrated significant reductions in computation time for both training and testing stages.
• Outperformed the $L_{2,1}$-based algorithm and standard machine learning approaches in efficiency.

Conclusions:

• The proposed framework effectively integrates nondifferentiable functions into random-weight neural networks.
• This approach offers a computationally efficient alternative without sacrificing predictive performance.
• The findings suggest a promising direction for optimizing machine learning algorithms.