Complex-valued soft-log threshold reweighting for sparsity of complex-valued convolutional neural networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Complex-valued convolutional neural networks (CVCNNs) excel at classifying complex signals and synthetic aperture radar (SAR) data.
• CVCNNs often suffer from redundancy and heavy computational costs due to complex-valued parameters.
• Model sparsity is a key technique to mitigate redundancy without significant performance degradation.

Purpose of the Study:

• To introduce a novel algorithm, complex-valued soft-log threshold reweighting (CV-SLTR), for designing sparse CVCNNs.
• To reduce the number of weight parameters and simplify the structure of CVCNNs.
• To address the limited research on sparsity in CVCNNs.

Main Methods:

• Developed a complex-valued log-sum threshold method, considering the unique properties of complex numbers.
• Created distinct complex-valued soft and log-sum thresholding techniques for pruning weights in complex-valued convolutional (CConv) and complex-valued fully connected (CFC) layers.
• Optimized sparsity thresholds during backward propagation using a sparsity budget and demonstrated convergence with stochastic gradient descent (SGD).

Main Results:

• The CV-SLTR algorithm effectively achieves sparsity in CVCNNs.
• Experiments on RadioML 2016.10A and S1SLC-CVDL datasets confirmed the algorithm's efficiency.
• The proposed method demonstrates fast sparsity achievement while maintaining high classification accuracy.

Conclusions:

• The CV-SLTR algorithm is a feasible and promising approach for creating sparse CVCNNs.
• This technique offers a way to reduce computational complexity and improve efficiency in complex signal processing tasks.
• The findings highlight the potential of CV-SLTR for practical applications requiring efficient CVCNNs.