Adaptive parallel decision deep neural network for high-speed equalization.

Area of Science:

• Optical Communications
• Digital Signal Processing
• Machine Learning

Background:

• High-speed optical wire-line transmission relies heavily on effective equalization.
• Traditional equalizers can face processing speed limitations due to feedback path timing constraints.
• Deep Neural Networks (DNNs) offer a promising digital signal processing approach for feedback-free signaling.

Purpose of the Study:

• To propose a resource-efficient DNN equalizer architecture for high-speed optical transmission.
• To investigate a parallel decision DNN that reduces hardware complexity.
• To evaluate the performance and training convergence of the proposed equalizer.

Main Methods:

• A parallel decision DNN architecture was developed, replacing the soft-max layer with a hard decision layer.
• This design enables multi-symbol processing within a single neural network.
• The neuron increment scales linearly with layer count, unlike duplication methods.

Main Results:

• The proposed parallel decision DNN demonstrates competitive performance against traditional equalizers (e.g., 15-tap feed forward equalizer with 2-tap decision feedback equalizer).
• Achieved performance at 28GBd and 56GBd with four-level pulse amplitude modulation and 30dB loss.
• Exhibited significantly faster training convergence compared to conventional counterparts.
• An adaptive mechanism based on forward error correction was explored.

Conclusions:

• The parallel decision DNN equalizer is an efficient architecture for high-speed optical transmission.
• It offers a favorable trade-off between performance, hardware resources, and training speed.
• This approach advances feedback-free signaling in optical communication systems.