Machine Learning and Neuromorphic Computing

About

The Machine Learning and Neuromorphic Computing research group is driving innovation at the intersection of photonics and artificial intelligence. Our mission is to harness the transformative potential of machine learning to enhance and reimagine photonic technologies, spanning across optical communication, system optimization, and hardware design.

Our research is focused across three interconnected areas:

We first develop computationally efficient digital algorithms tailored to correct channel impairments in optical transmission systems. These efforts combine theoretical models with real-world implementation, aiming to boost the reliability and efficiency of modern optical networks.

Secondly, we are advancing intelligent optimization tools that streamline the design and deployment of photonic devices and networks. By applying state-of-the-art machine learning techniques, we reduce design complexity and costs while improving system performance.

The third core area involves the development of novel photonic components that embed machine learning capabilities directly into the optical domain. These neuromorphic systems promise to significantly increase computational speed and energy efficiency—key to meeting the demands of next-generation data processing.

Together, our work paves the way for smarter, faster, and more efficient photonic systems with broad implications across communications, sensing, and computing.

Our Projects

UK Multidisciplinary Centre for Neuromorphic Computing

The UK Multidisciplinary Centre for Neuromorphic Computing is pioneering brain-inspired, energy-efficient computing technologies to address the sustainability challenges facing today’s digital infrastructure and artificial intelligence systems.

Led by the Aston Institute of Photonic Technologies (AiPT) at Aston University, and supported by EPSRC, the centre brings together world-leading researchers from across disciplines - including neuroscience, non-conventional computing algorithms, photonics, opto- and nano-electronics, materials science, and device engineering - to transform the way computing systems are designed and used.

The Centre team includes researchers from Aston University, University of Oxford, University of Cambridge, University of Southampton, Queen Mary University of London, Loughborough University, University of Strathclyde. The Centre focus on impact is supported by a broad network of industry partners, including Microsoft Research, Thales, BT, QinetiQ, Nokia Bell Labs, Hewlett Packard Labs, Leonardo, Northrop Grumman, and a number of SMEs.

Find out more - UK Multidisciplinary Centre for Neuromorphic Computing | Aston University

Reducing Neural Network Complexity for Optical Channel Equalization

This project explores advanced neural network compression techniques to reduce computational complexity in optical communication systems. By applying quantization, weight clustering, and pruning to feed-forward and recurrent neural network equalizers, the team evaluates the trade-offs between performance and efficiency. Using both simulated and experimental data, results show that compressed neural networks can outperform traditional digital back-propagation equalizers while maintaining comparable complexity. The work also introduces new benchmarking metrics and highlights future directions for achieving high-performance, energy-efficient optical equalization.

P. J. Freire et al., "Reducing Computational Complexity of Neural Networks in Optical Channel Equalization: From Concepts to Implementation," Journal of Lightwave Technology, vol. 41, no. 14, pp. 4557-4581, 15 July15, 2023, doi: 10.1109/JLT.2023.3234327

P. Freire et al., "Computational Complexity Optimization of Neural Network-Based Equalizers in Digital Signal Processing: A Comprehensive Approach," in Journal of Lightwave Technology, doi: 10.1109/JLT.2024.3386886.

Parallelization of Recurrent Neural Network-Based Equalizer for Coherent Optical Systems via Knowledge Distillation

This project explores a novel approach to reduce the computational complexity of recurrent neural network (RNN)-based equalizers in coherent optical systems. By applying a knowledge distillation (KD) technique, the performance of a bidirectional long short-term memory (biLSTM) network is transferred to a more computationally efficient 1D convolutional neural network (1D-CNN). The new feed-forward architecture enables parallelization, significantly reducing inference time while maintaining strong equalization performance. Evaluations using both simulated and experimental data show that the distilled 1D-CNN achieves comparable Q-factors to the original biLSTM, offering a promising solution for real-time optical signal processing.

S. Srivallapanondh et al., "Parallelization of Recurrent Neural Network-Based Equalizer for Coherent Optical Systems via Knowledge Distillation," in Journal of Lightwave Technology, vol. 42, no. 7, pp. 2275-2284, 1 April1, 2024, doi: 10.1109/JLT.2023.3337604

Artificial Neural Networks for Photonics: From Algorithms to Implementation

This tutorial–review bridges the gap between artificial neural network (ANN) algorithms and their practical implementation in photonic systems. Aimed at both optical and computer science communities, it explores the design and optimization of neural network models tailored for photonic applications. The review highlights key ANN architectures relevant to photonics, linking theoretical concepts to physical realizations in hardware. It also covers recent advances in applying ANNs across optical communications, imaging, sensing, and laser/material design. A key focus is placed on evaluating model complexity and exploring compression techniques that support the transition from algorithm development to efficient optical hardware integration.

P. Freire, E. Manuylovich, J. E. Prilepsky, and S. K. Turitsyn, "Artificial neural networks for photonic applications—from algorithms to implementation: tutorial," Adv. Opt. Photon. 15, 739-834 (2023).

Time and Frequency Domain Learned MIMO Volterra Equalizers

This project introduces advanced multi-channel nonlinear equalizers (NLEs) based on machine learning and Volterra series transfer functions to enhance optical communication system capacity. Initially, a time-domain MIMO equalizer was developed using efficient FIR filter reuse to reduce computational complexity while maintaining high performance. The work then evolved into a hybrid approach, where linear components were processed in the frequency domain, further optimizing performance through block-based processing. Extensive numerical evaluations demonstrated a 5×5 MIMO configuration achieving over 2dB effective SNR improvement compared to single-channel equalizers, establishing a foundation for future high-capacity WDM systems.

N. Castro, S. Boscolo. A. D. Ellis. S. Sygletos, “Design of Time-Domain Learned Volterra Equaliser for WDM Systems,” 28th Internaltional Conference on Optical Network Design and Modelling, 06-09 May 2024, Madrid Spain

N. Castro, S. Boscolo, A. D. Ellis, S. Sygletos,” Design Aspects of Frequency-Domain Learned MIMO Volterra Equalisers,” CLEO 2024

Improving Robustness of Analog Neural Networks with Explainable Regularization

This project addresses the noise sensitivity challenge in analog neural networks—an emerging energy-efficient alternative to digital computing. By analyzing the effects of additive noise on feed-forward architectures, the team developed a novel, noise-agnostic regularization strategy that improves model robustness without relying on specific hardware configurations. Supported by theoretical insights and validated through computer vision tasks, the approach achieved over 53% accuracy improvement in noisy conditions, marking a significant step toward more reliable analog AI systems.

A. Duque, P. Freire, E. Manuylovich. D. Stoliarov, J. Prilepsky, S. Turitsyn “Improving Analog Neural Network Robustness: A Noise-Agnostic Approach with Explainable Regularizations” Second Workshop on Machine Learning with New Compute Paradigms at NeurIPS 2024.