Photonic neuromorphic computing with recurrent optical spectrum slicing neural networks (ROSS-NN)

We report on a novel recurrent photonic node for use at photonic recurrent neural networks and reservoir computing architectures. Based on this node, we propose a format agnostic receiver-equalizer for the next generation high-baudrate optical links.
Photonic neuromorphic computing with recurrent optical spectrum slicing neural networks (ROSS-NN)
Like

Neuromorphic computing aims to mimic the way the human brain works in order to solve problems in which conventional computing struggles. Photonic computing is a promising route towards ultrafast processing while maintaining low power consumption. The merge of these two fields, that is photonic neuromorphic computing, has found applications in a wide area of problems, with imaging and optical communications being the most significant. The key in the aforementioned problems is that the information is primarily in an optical form, thus photonic computing arises as a natural extension. Nowadays, optical communications enable the highways which deliver in an ultra-fast manner data, one of most important resources in the era of digital transition. Pushed to deliver the highest rate of data in every corner of the planet, optical communications take advantage of sophisticated digital signal processing for data recovery, consuming significant amount of power. Reducing the power consumption, while recovering the ever-increasing distortions of the transmitted optical signals is of paramount importance for the next generation optical links.

During the last years, our group has experimented with unconventional-neuromorphic techniques for the mitigation of transmission impairments in an energy-efficient manner. Various classes of recurrent neural networks like Long-Short Term Memory (LSTM) and reservoir computing (RC) have been investigated and compared either for the short-reach or for the long-haul optical channels. Reservoir computer refers to a recurrent neural network with untrained reservoir weights, while the training is taking place only in the readout layer, thus simplifying and accelerating the whole procedure. For the short-reach intensity modulation/direct detection (IM/DD) links which are more sensitive to power consumption, photonic nodes and photonic RC networks are good candidates. In our efforts to improve the efficiency of photonic recurrent networks, we examined splitting a reservoir of photonic nodes in smaller sub-reservoirs, a strategy that had already shown efficiency 1. Another promising idea was given by Ranzini et al.2, in which optical filters spectrally slice the incoming signal before reservoir processing. So, we configured one of our previous photonic RC works 3 by splitting it in sub-RCs and we configured them to slice spectrally the input. The performance of this system significantly outperforms state-of-the-art digital algorithms. After thorough investigation, we realized that even single-node sub-RCs (in practice single recurrent nodes) offer improved compensation performance. Eventually, in order to offer improved error rate performance and preserve the receiver complexity at moderate levels, we end up with the proposed scheme consisting of only two nodes performing recurrent optical spectrum slicing (Figure 1). 

Figure 1. The proposed recurrent filter node, a network of such nodes and a two-node telecom receiver.

In signals affected by chromatic dispersion, direct optical frequency processing is the key for equalization. By multiplying the transfer function of the dispersive transmission with the versatile ROSS transfer functions, the frequency nulls due to power fading can be relaxed, while applying frequency diversity (Figure 2), the power fading problem can be practically eliminated. ROSS nodes operate as data rate agnostic, passive equalizers, thus they can scale up to hundreds of Gbaud. Moreover, they can also operate as format agnostic equalizers. This is attributed to the coherent nature of the filter nodes, which receive and process directly the input light prior to photodetection. In this way, quadrature amplitude modulation formats can be also equalized, and most importantly, in a self-coherent way, without the need for a cumbersome coherent receiver. So, using the same scheme at the receiver we can simultaneously achieve dispersion management in IM/DD systems, reception and equalization in coherent systems. Last but not least, the aforementioned functionalities are accompanied by bandwidth relaxation in the photodiodes and analog-to-digital converters. This is owed to the spectral slicing that distributes the overall signal bandwidth to two, three or more detectors, reducing the bandwidth requirements by at least 20%, with subsequent consumption gains especially in the digital domain. 

Figure 2. The effect of the recurrent node to the signal spectrum reduces the spectral dips due to power fading effect. Two recurrent nodes act complementary, further unveiling the diminished frequencies.

By accelerating consuming digital algorithms and by reducing the required bandwidth of critical components, the ROSS receiver could constitute a viable and attractive alternative in the typical coherent receiver that aims to conquer the full range of optical interconnects. For a rough estimate, a typical 4 wavelength IM/DD pluggable operating at 800 Gbps would consume 2 W more than the proposed ROSS receiver, while the coherent alternatives would be even more power hungry. However, we do not want to stop here. We examine the capability of ROSS networks as convolutional accelerators for image processing 4, as sensing systems and even as physical unclonable functions 5. Maybe two or more combinations of the aforementioned operations could be functional simultaneously due to the inherent parallelism of the photonic technology. Furthermore, the experimental evaluation of the scheme in reconfigurable photonic platforms, along with our partners in HORIZON 2020 projects, will be a significant milestone for us.

 

  1. Freiberger, M. et al. Improving Time Series Recognition and Prediction With Networks and Ensembles of Passive Photonic Reservoirs. IEEE Journal of Selected Topics in Quantum Electronics 26, 1–11 (2020).
  2. Ranzini, S. M., Dischler, R., da Ros, F., Bulow, H. & Zibar, D. Experimental Investigation of Optoelectronic Receiver With Reservoir Computing in Short Reach Optical Fiber Communications. Journal of Lightwave Technology 39, 2460–2467 (2021).
  3. Mesaritakis, C., Sozos, K., Dermanis, D. & Bogris, A. Spatial Photonic Reservoir Computing based on Non-Linear Phase-to-Amplitude Conversion in Micro-Ring Resonators. in Optical Fiber Communication Conference (OFC) 2021 Tu1H.2 (OSA, 2021). doi:10.1364/OFC.2021.Tu1H.2.
  4. Tsirigotis, A., Tsilikas, I., Sozos, K., Bogris, A. & Mesaritakis, C. Filter-based photonic reservoir computing as a key-enabling platform for all-optical, high-speed processing of time-stretched images and telecomm data. in AI and Optical Data Sciences III (eds. Kitayama, K. & Jalali, B.) 50 (SPIE, 2022). doi:10.1117/12.2607438.
  5. Dermanis, D., Bogris, A., Rizomiliotis, P. & Mesaritakis, C. Photonic Physical Unclonable Function based on Integrated Neuromorphic Devices. Journal of Lightwave Technology (2022) doi:10.1109/JLT.2022.3200307.

Please sign in or register for FREE

If you are a registered user on Research Communities by Springer Nature, please sign in

Subscribe to the Topic

Electrical and Electronic Engineering
Technology and Engineering > Electrical and Electronic Engineering

Related Collections

With collections, you can get published faster and increase your visibility.

Thermal Engineering for Sustainability

This collection will publish papers on the topic of thermal engineering with a particular focus on applications related to sustainability.

Publishing Model: Open Access

Deadline: Mar 30, 2024

Consumer waste valorization

This collection will focus on chemical, materials, mechanical and other engineering advances tackling post-consumer solid waste streams such as textiles and paper, food, electronics and plastics.

Publishing Model: Open Access

Deadline: May 31, 2024