[Paper Review] Spiking Neural Network Equalization on Neuromorphic Hardware for IM/DD Optical Communication

Elias Arnold, Georg Böcherer+5 more|arXiv (Cornell University)|Jun 1, 2022

Advanced Memory and Neural Computing4 citations

TL;DR

This paper proposes a spiking neural network (SNN) equalizer/demapper implemented on the BrainScaleS-2 neuromorphic hardware for intensity-modulation/direct-detection (IM/DD) optical communication. The SNN achieves a bit error rate (BER) of $2 \times 10^{-3}$ with a hardware penalty of less than 1 dB, outperforming software-simulated linear minimum mean square error (LMMSE) equalizers and demonstrating the feasibility of analog neuromorphic processing for low-power optical signal detection at 200 Gbit/s with 12% FEC overhead.

ABSTRACT

A spiking neural network (SNN) non-linear equalizer model is implemented on the mixed-signal neuromorphic hardware system BrainScaleS-2 and evaluated for an IM/DD link. The BER 2e-3 is achieved with a hardware penalty less than 1 dB, outperforming numeric linear equalization.

Motivation & Objective

• To evaluate the feasibility of spiking neural networks (SNNs) for nonlinear equalization in IM/DD optical communication systems using neuromorphic hardware.
• To demonstrate that SNN-based equalization on analog neuromorphic hardware can achieve performance comparable to software-simulated SNNs and surpass traditional linear equalizers.
• To validate the reliability and low hardware penalty of in-the-loop trained SNNs on mixed-signal neuromorphic systems like BrainScaleS-2 for real-time optical signal processing.
• To explore the potential of neuromorphic hardware to replace power-hungry digital signal processing in optical transceivers, particularly by eliminating the need for high-speed ADCs.

Proposed method

• An SNN with one hidden layer of 40 leaky-integrate-and-fire (LIF) neurons and four leaky-integrate (LI) readout neurons is designed for joint equalization and demapping of PAM4 signals.
• The SNN is trained using in-the-loop (ITL) learning with surrogate gradients, where the forward pass is executed on the BrainScaleS-2 hardware and weight updates are computed on a host computer.
• The system uses mixed-signal neuromorphic hardware with 6-bit configurable synaptic weights, analog membrane potential integration, and digital spike communication to enable low-power, continuous-time processing.
• The SNN processes 200 Gbit/s IM/DD signals over 4 km of fiber in the O-band, with chromatic dispersion and additive white Gaussian noise (AWGN) modeled in simulation.
• Hardware experiments are conducted using the BSS-2 system, with input signals preprocessed via root-raised cosine (RRC) filtering and photodiode squaring at the receiver.
• Network performance is evaluated by comparing BER results between simulated SNNs, software ANNs, and hardware-implemented SNNs on BSS-2, with decision boundaries optimized for FEC threshold of $2 \times 10^{-3}$.

Experimental results

Research questions

• RQ1 Can a spiking neural network implemented on neuromorphic hardware achieve BER performance comparable to software-simulated SNNs in an IM/DD optical link?
• RQ2 Does neuromorphic SNN processing outperform conventional linear equalization (LMMSE) in terms of BER for PAM4 signals impaired by chromatic dispersion and noise?
• RQ3 What is the hardware penalty of deploying an SNN on BrainScaleS-2 compared to software simulation, and can it be kept below 1 dB at a target BER of $2 \times 10^{-3}$?
• RQ4 Can in-the-loop training on analog neuromorphic hardware enable reliable, low-power signal processing suitable for real-time optical transceivers?
• RQ5 Can SNN-based equalization eliminate the need for high-power analog-to-digital conversion in optical receivers by processing spikes directly from the photodiode?

Key findings

• The SNN equalizer implemented on BrainScaleS-2 achieves a BER of $2 \times 10^{-3}$ with a hardware penalty of less than 1 dB compared to the software-simulated SNN.
• The hardware-implemented SNN outperforms a 7-tap LMMSE linear equalizer, demonstrating the advantage of nonlinear processing in signal recovery.
• The SNN trained on BSS-2 shows superior BER performance compared to both single- and two-hidden-layer ReLU-based artificial neural networks (ANNs) in simulation.
• Membrane potential traces and spike activity on BSS-2 confirm that the correct readout neuron is selectively activated, indicating confident decision-making.
• Weight matrices learned on BSS-2 show that input weights from the central tap dominate, reflecting their critical role in classifying the current symbol.
• The system processes one input sample in 30 μs, with potential for parallelization to meet 200 Gbit/s throughput requirements.