Diego Arguello Ron (ASTON)

Diego’s research focused on the implementation and improvement of neural network-based equalizers in optical communication systems, specifically targeting the reduction of computational complexity and enhancing noise resistance. He explored novel approaches involving pruning techniques and optical phase conjugation to address challenges such as signal impairment over long distances and the high computational load typically associated with neural network equalizers. Key outcomes include demonstrating the potential for significant memory and complexity reductions without a substantial loss in performance, offering a viable path toward the deployment of advanced, efficient optical communication technologies. Combining machine learning with optical phase conjugation creates a powerful tool for making optical communication networks better at handling noise and signal problems. This approach mixes optical and computer methods to make the systems more robust and capable of sending data further and more reliably.

Figure 10: Structure of a communication channel that is equalized using a pruned and quantized neural network deployed on resource-restricted hardware (Raspberry Pi 4 or Nvidia Jetson Nano)

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 still 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 dispersion-induced memory. To implement the NN-based optical channel equalizer in hardware, a substantial complexity reduction is needed, while we have to keep an acceptable performance level of the simplified NN model.

Figure 11: (a) Q-factor achieved for pruned and quantized models versus the level of sparsity for datasets corresponding to three launch powers: 0 dBm, 1 dBm, and 2 dBm; The solid lines correspond to the Q-factor achieved by the original model. The dashed lines show the Q-factor when only linear equalization (LE) is implemented. (b) Q-factor achieved after pruning compared to the one achieved after both pruning and quantization, for different levels of sparsity and for a dataset corresponding to the 1 dBm launch power. The blue and red solid lines correspond to the Q-factor achieved by the original model and the one achieved by this model after quantization, respectively

Diego, together with his coauthors in [1], addressed the complexity reduction problem by applying pruning and quantization techniques to an NN-based optical channel equalizer. They used an exemplary NN architecture, the multi-layer perceptron (MLP), to mitigate the impairments for 30 GBd 1000 km transmission over a standard single-mode fiber, and demonstrate that it is feasible to reduce the equalizer’s memory byup to 87.12%, and its complexity by up to 78.34%, without noticeable performance degradation.

Figure 12: Energy consumption for Raspberry Pi 4 and Nvidia Jetson Nano. The blue section represents the energy consumption per recovered symbol when using the compressed model, and its relative energy cost is expressed as a percentage with respect to the sum of the energy consumed by both the original and compressed models. Likewise, the red section describes the energy consumption per recovered symbol when using the original model and its relative energy cost.

Additionally, they accurately define the computational complexity of a compressed NN-based equalizer in the digital signal processing (DSP) sense, examine the impact of using hardware with different CPU and GPU features on the power consumption and latency for the compressed equalizer and verify the developed technique experimentally, by implementing the reduced NN equalizer on two standard edge-computing hardware units: Raspberry Pi 4 and Nvidia Jetson Nano, which are used to process the data generated via simulating the signal’s propagation down the optical-fiber. These results have been published in [1].

Diego and his co-authors in [2] also introduced a novel extended methodology for developing low-complexity neural network (NN) based equalizers to address impairments in high-speed coherent optical transmission systems. They present a comprehensive exploration and comparison of deep model compression techniques applied to feed-forward and recurrent NN designs, assessing their impact on equalizer performance. Our investigation encompasses quantization, weight clustering, pruning, and other cutting-edge compression strategies. Diego and his co-authors propose and evaluate a Bayesian optimization-assisted compression approach that optimizes hyperparameters to simultaneously enhance performance and reduce complexity, and introduce four distinct metrics (RMpS, BoP, NABS, and NLGs) to quantify computing complexity in various compression algorithms. These metrics serve as benchmarks for evaluating the relative effectiveness of NN equalizers when compression approaches are employed. The analysis is completed by evaluating the trade-off between compression complexity and performance using simulated and experimental data. By employing optimal compression techniques, and published in [2], they demonstrate the feasibility of designing a simplified NN-based equalizer surpassing the performance of conventional digital back-propagation (DBP) equalizers with only one step per span. This is achieved by reducing the number of multipliers through weighted clustering and pruning algorithms, with results having been published in [2].

Figure 13: Performance (Q-factor improvement) of the Post Training Quantization (Heterogeneous Approach) for the simulated (Sim1 and Sim2) and experimental (Exp) transmission datasets for different quantisation methods applied for neural network weights.

Furthermore, Diego and co-authors in [2] highlight that an NN-based equalizer can achieve better performance than the full electronic chromatic dispersion compensation block while maintaining a similar level of complexity.

Key publications by Diego A. Ron:

[1] Ron, D. A., Freire, P. J., Prilepsky, J. E., Kamalian-Kopae, M., Napoli, A., & Turitsyn, S. K. (2022). Experimental implementation of a neural network optical channel equalizer in restricted hardware using pruning and quantization. Scientific Reports, 12(1), 8713.

[2 Freire PJ, Napoli A, Spinnler B, Anderson M, Ron DA, Schairer W, Bex T, Costa N, Turitsyn SK, Prilepsky JE. Reducing computational complexity of neural networks in optical channel equalization: From concepts to implementation. Journal of Lightwave Technology. 2023 Jan 5;41(14):4557-81

[3] Ron, D. A., Nurlybayeva, K., Kamalian-Kopae, M., Ali, A. A., Turitsyna, E., & Turitsyn, S. (2022, September). On the Impact of the Optical Phase Conjugation on the Computational Complexity of Neural Network-Based Equalisers. In European Conference and Exhibition on Optical Communication (pp. We5-29). Optica Publishing Group.

[4] Ron, D. A., Kamalian-Kopae, M., & Turitsyn, S. (2021, June). Noise-Resistant Optical Implementation of Analogue Neural Networks. In European Quantum Electronics Conference (p. jsiv_3_2). Optica Publishing Group.

[5] Ron, D. A. (2023, May). Combination of Optical Phase Conjugation and Advanced Pruning Techniques to reduce the Computational Complexity of Neural Network-Based Equalisers. In CLEO: Science and Innovations (pp. SM3I-6). Optica Publishing Group.

[6] Nurlybayeva, K., Ron, D. A., Kamalian-Kopae, M., Turitsyna, E., & Turitsyn, S. (2022, October). Noise-Resistant Crowd Equalisation for Optical Communication Systems Based on Machine Learning. In Frontiers in Optics (pp. FM3D-2). Optica Publishing Group.

[7] Nurlybayeva, K., Ron, D. A., Kamalian-Kopae, M., Turitsyna, E., & Turitsyn, S. (2022, November). Implementation of Noise-Resistant Crowd Equalisation in Optical Communication Systems with Machine Learning DSP. In 2022 Asia Communications and Photonics Conference (ACP) (pp. 753-756). IEEE.