Research on improved convolutional wavelet neural network.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Traditional artificial neural networks (ANN), including Backpropagation Neural Networks (BPNN), Radial Basis Function Neural Networks (RBFNN), and Convolutional Neural Networks (CNN), suffer from issues like local minima, instability, and limited precision.
• Improvements in training speed and accuracy are crucial for advancing these models.

Purpose of the Study:

• To develop novel neural network architectures that overcome the limitations of existing models.
• To enhance the performance, particularly precision and training speed, of deep learning models.

Main Methods:

• Wavelet Neural Network (WNN) was developed using BPNN structure and wavelet functions as activation functions to improve upon BPNN and RBFNN.
• A WNN-based Convolutional Wavelet Neural Network (CWNN) was proposed, replacing CNN's fully connected layers with WNN.
• A Wavelet-based Convolutional Neural Network (WCNN) was introduced, incorporating wavelet transformation as activation functions within CNN's convolutional pooling layers.

Main Results:

• WNN demonstrated superior performance compared to BPNN and RBFNN, effectively solving their inherent problems.
• CWNN achieved lower mean square error and error rates than standard CNN, indicating enhanced maximum precision.
• WCNN further improved upon CWNN by reducing mean square error and error rates, signifying superior maximum precision.

Conclusions:

• The proposed WNN, CWNN, and WCNN architectures offer significant performance improvements over traditional neural networks.
• Wavelet functions and transformations are effective in enhancing the precision and stability of deep learning models.
• These novel architectures provide a promising direction for future research in artificial neural networks.