On Training Neural Network Decoders of Rate Compatible Polar Codes via Transfer Learning.

Area of Science:

• Coding Theory
• Machine Learning
• Telecommunications

Background:

• Rate-compatible polar codes are crucial for modern wireless communication standards like 5G.
• Training neural network decoders (NNDs) for these codes can be computationally intensive.
• Existing methods may struggle with underfitting in NNDs due to model complexity limitations.

Purpose of the Study:

• To propose an efficient transfer learning technique for training NNDs for rate-compatible polar codes.
• To leverage the inclusion property of rate-compatible polar codes for improved training.
• To address the underfitting problem in NND training.

Main Methods:

• A family of rate-compatible polar codes based on 5G new radio definitions was used.
• A transfer learning approach was developed, initializing NND training for a higher rate code with weights from a lower rate code.
• Numerical results were used to compare training times and performance against separate learning.

Main Results:

• The proposed transfer learning method significantly accelerates the training of NNDs for rate-compatible polar codes.
• Transfer learning effectively resolves the underfitting problem often encountered with low-complexity NND models.
• Trained NNDs for lower rate codes serve as effective initializations for training NNDs of subsequent, higher rate codes.

Conclusions:

• Transfer learning offers a more efficient and effective approach to training NNDs for rate-compatible polar codes.
• This technique enhances training speed and mitigates underfitting, improving decoder performance.
• The findings have direct implications for optimizing 5G communication systems and future wireless technologies.