Minimalist Deployment of Neural Network Equalizers in a Bandwidth-Limited Optical Wireless Communication System with...

Area of Science:

• Optical communication systems
• Machine learning for signal processing
• Deep learning architectures

Background:

• Recurrent neural networks (RNNs), particularly bidirectional gated recurrent units (biGRUs), excel at handling nonlinear damage and inter-symbol interference (ISI) in optical communications.
• However, the inherent recursive structure of biGRUs limits computational parallelization, leading to lower equalization rates.
• There is a need for faster equalization methods without sacrificing performance in optical communication systems.

Purpose of the Study:

• To propose and evaluate a minimalist 1D convolutional neural network (CNN) equalizer.
• To improve equalization speed in optical communication systems by employing knowledge distillation (KD) from a biGRU model.
• To investigate the effectiveness of KD in regression problems for neural network model compression and acceleration.

Main Methods:

• Knowledge distillation (KD) was applied to train a 1D CNN equalizer using a pre-trained biGRU model as the teacher.
• The performance of the KD-trained 1D CNN was compared against the original biGRU and a 1D CNN trained without KD.
• Key performance metrics included the Q-factor (signal quality) and equalization velocity (speed).

Main Results:

• The KD-trained 1D CNN achieved a 1 dB increase in Q-factor compared to a 1D CNN without KD.
• KD improved the received optical power (RoP) sensitivity of the 1D CNN by 0.89 dB.
• The 1D CNN equalizer demonstrated a 97% reduction in computational time and a 99.3% reduction in trainable parameters compared to the biGRU, with only a 0.5 dB Q-factor penalty.

Conclusions:

• A minimalist 1D CNN equalizer, optimized via KD from a biGRU, offers a promising solution for high-speed optical communication systems.
• This approach significantly accelerates equalization while preserving high signal quality, making it suitable for practical deployment.
• The study highlights the efficacy of KD for model compression and performance enhancement in regression tasks within optical communication signal processing.