Data classification based on fractional order gradient descent with momentum for RBF neural network

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neural Networks

Background:

• Weight-updating methods are crucial for neural network performance.
• Radial basis function (RBF) neural networks can suffer from training oscillations.
• Efficient training algorithms are needed for complex data classification tasks.

Purpose of the Study:

• To propose a novel weight-updating method for RBF neural networks to address training oscillations.
• To enhance the convergence speed and overall performance of RBF neural network training.
• To validate the proposed method's effectiveness in data classification.

Main Methods:

• A fractional order gradient descent with momentum for RBF neural network (FOGDM-RBF) weight updating is proposed.
• An adaptive learning rate is incorporated to accelerate convergence.
• The algorithm is tested on the Iris and MNIST datasets for data classification.

Main Results:

• The FOGDM-RBF algorithm demonstrates monotonicity and convergence, verifying theoretical results.
• Non-parametric statistical tests (Friedman, Quade) confirm superior performance compared to other algorithms.
• Analysis of fractional order, learning rate, and batch size influences performance.

Conclusions:

• The proposed FOGDM-RBF algorithm effectively accelerates gradient descent convergence.
• The method significantly improves RBF neural network performance, achieving high accuracy and validity in data classification.
• This approach offers a robust solution for mitigating oscillations in RBF neural network training.