Robust Reconstructed Neural Network With Spectral Reshaping Activation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Neural Networks (NNs) are powerful information processing models.
• NNs often suffer from incorrect neuron activation due to noise.
• Existing methods struggle with the boundary effects of compound noises.

Purpose of the Study:

• To propose a Robust Reconstructed Neural Network (RRNN) for enhanced performance in noisy environments.
• To introduce Spectral Reshaping Activation (SRA) as a novel activation function for NNs.
• To develop a Hierarchical Gradient Descent (HGD) algorithm for RRNN parameter optimization.

Main Methods:

• Designed Spectral Reshaping Activation (SRA) to shrink noise spectrums via spectral subtraction.
• Developed Hierarchical Gradient Descent (HGD) for RRNN parameter updates.
• Incorporated a noise-contrastive degree in the loss function for robustness.

Main Results:

• SRA effectively reshapes noise space for easier coverage by RRNN.
• RRNN demonstrates robust performance across different noise types.
• Theoretical validation confirms the robustness of RRNN.

Conclusions:

• The proposed RRNN with SRA significantly improves robustness against noisy samples.
• RRNN outperforms existing methods in handling noisy data.
• The developed HGD algorithm ensures effective parameter optimization for robust performance.