Experimental Implementation of a Neural Network Optical Channel Equalizer in Restricted Hardware Using Pruning and Quantization

Abstract

The deployment of artificial neural networks-based optical channel equalizers on edge-computing devices is critically important for the next generation of optical communication systems. However, this is a highly challenging problem, mainly due to the computational complexity of the artificial neural networks (NNs) required for the efficient equalization of nonlinear optical channels with large memory. To implement the NN-based optical channel equalizer in hardware, a substantial complexity reduction is needed, while keeping an acceptable performance level. In this work, we address this problem by applying pruning and quantization techniques to an NN-based optical channel equalizer. We use an exemplary NN architecture, the multi-layer perceptron (MLP), and address its complexity reduction for the 30 GBd 1000 km transmission over a standard single-mode fiber. We demonstrate that it is feasible to reduce the equalizer’s memory by up to 87.12%, and its complexity by up to 91.5%, without noticeable performance degradation. In addition to this, we accurately define the computational complexity of a compressed NN-based equalizer in the digital signal processing (DSP) sense and examine the impact of using different CPU and GPU settings on power consumption and latency for the compressed equalizer. We also verify the developed technique experimentally, using two standard edge-computing hardware units: Raspberry Pi 4 and Nvidia Jetson Nano.

Implementation of Noise-Resistant Crowd Equalisation in Optical Communication Systems with Machine Learning DSP

“…As is shown, the sensitivity of the performance of these equalisers decreases as the size of the crowd increases. On the other hand, CC is defined as the number of complex-valued multiplications [3]:…”

“…Although the performance of some of these digital signal processing (DSP) techniques is impressive in restoring the original state of the received signal, these equalisers often require complex architecture with complicated elements such as long shortterm memory nodes [1,2]. This increases the computational complexity (CC) of such solutions beyond what is acceptable for real-time implementation, especially when compared to alternatives such as digital backpropagation, and at odds with the promise of low footprint solutions [3]. Moreover, their architecture is fundamentally different from the one that digital computers have, making this hardware not suitable for them.…”

A Review of AIoT-based Edge Devices and Lightweight Deployment

“…Usually, we first apply knowledge distillation to improve the performance, then apply network pruning [132], [135] or quantization. Or first, apply network pruning, then quantization [88], [90], [95], [96], [100]. In summary, knowledge distillation should be applied first to ensure performance; then network pruning for model compression; quantization should be placed in the last step because the training of the quantized network is challenging.…”

Reducing Computational Complexity of Neural Networks in Optical Channel Equalization: From Concepts to Implementation

“…According to Ref. [39], [40], the compression can often be accomplished with little loss of accuracy and, in some situations, the accuracy may even rise [41]. Three methods of network compression are discussed below: pruning, weight clustering, and quantization.…”

“…In this study, to validate the robustness of such a NN equalizer on the quantization, they reduced the bit precision of weights to up to 2 bits, observing mostly only minor performance degradation. Finally, in [41], an MLP equalizer was used to mitigate the impairments in a 30 GBd 1000 km system. In this case, the PTQ strategy together with the traditional uniform 8 bits quantization was demonstrated using low-performing hardware (Raspberry Pi and Jetson nano).…”