Deep Sparse Learning for Automatic Modulation Classification Using Recurrent Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• Deep learning models, particularly Recurrent Neural Networks (RNNs), show promise in Automatic Modulation Classification (AMC).
• Overparameterization in deep neural networks leads to large model sizes, hindering deployment in resource-constrained environments like the Internet of Things (IoT).
• Sparse learning offers a solution to reduce model complexity by removing redundant parameters without performance degradation.

Purpose of the Study:

• To propose a novel sparse learning algorithm for training sparsely connected neural networks directly.
• To reduce the parameter count of deep learning models for applications like AMC, addressing deployment challenges.
• To investigate the impact of sparsity on network performance and generalization ability.

Main Methods:

• Developed a sparse learning algorithm leveraging weight magnitude and gradient momentum statistics.
• Validated the method on benchmark datasets (MNIST, CIFAR10).
• Applied the algorithm to RNNs for AMC, exploring various pruning strategies for recurrent and non-recurrent connections.

Main Results:

• The proposed method effectively reduced neural network parameters while preserving model performance.
• Experimental results on AMC tasks demonstrated the efficacy of the sparse learning approach.
• Appropriate levels of sparsity were found to improve the generalization ability of the networks.

Conclusions:

• The developed sparse learning algorithm is effective for parameter reduction in deep neural networks.
• The method successfully maintains or improves performance in Automatic Modulation Classification tasks.
• Sparse learning presents a viable strategy for efficient deep learning model deployment, especially in IoT contexts.