Source-linked AI summary
All-optical Reservoir Computing
François Duport, Bendix Schneider, Anteo Smerieri, Marc Haelterman, Serge Massar
TL;DR
The paper addresses whether Reservoir Computing can achieve strong information processing with an all-optical implementation. It builds a delay-loop reservoir from off-the-shelf telecommunications components using SOA saturation as the nonlinearity, and reports state-of-the-art capacities on several benchmark tasks. The main constraint is noise from SOA amplified spontaneous emission, which degrades some task and memory performance.
Problem
The paper investigates whether Reservoir Computing can be implemented all-optically using practical optical components while retaining strong benchmark performance.
Method
The authors build a single-node fiber delay-loop Reservoir Computer from off-the-shelf optical telecommunications components, using SOA saturation gain as its nonlinearity.
Results
The all-optical reservoir exhibits state-of-the-art capacities on several benchmark tasks.
Takeaways & Limitations
The experiment demonstrates the potential of optics for artificial-intelligence computing within the Reservoir Computing paradigm.
Takeaways & Limitations
SOA amplified spontaneous-emission noise lowers memory capacities and degrades performance on several tasks compared with the optoelectronic reservoir.
Abstract
from arXiv · showhide
Reservoir Computing is a novel computing paradigm which uses a nonlinear recurrent dynamical system to carry out information processing. Recent electronic and optoelectronic Reservoir Computers based on an architecture with a single nonlinear node and a delay loop have shown performance on standardized tasks comparable to state-of-the-art digital implementations. Here we report an all-optical implementation of a Reservoir Computer, made of off-the-shelf components for optical telecommunications. It uses the saturation of a semiconductor optical amplifier as nonlinearity. The present work shows that, within the Reservoir Computing paradigm, all-optical computing with state-of-the-art performance is possible.
1.Introduction
Reservoir Computing processes time-dependent inputs with a nonlinear dynamical system and has achieved strong performance in prior electronic and optoelectronic implementations. This work reports an experimental all-optical reservoir using a single nonlinear node, a delay loop, and SOA saturation.
- Reservoir Computing is a neural-network paradigm suited to time-dependent inputs, including speech recognition and time-series prediction.
- A Reservoir Computer uses a nonlinear dynamical reservoir driven by input and a trained linear output layer.
- Earlier electronic and optoelectronic Reservoir Computers achieved performance comparable to state-of-the-art digital implementations.
- Reservoir Computing is attractive for optics because it can use available nonlinearities directly instead of constructing logical gates and larger optical architectures.
- The paper reports the first all-optical experimental Reservoir Computer, using a fiber delay loop with one nonlinear node and SOA saturation gain.
- The experiment differs conceptually from previous implementations through its SOA-based nonlinearity and substantial spontaneous-emission noise.
2.The photonic hardware implementation
The reservoir uses a delayed-feedback architecture implemented with off-the-shelf optical telecommunications components. Its SOA supplies the nonlinearity, while desynchronized masking creates coupled internal variables and fading-memory dynamics.
- Reservoir-computing architecture: A reservoir computer processes an input signal through nonlinear recurrent node states and produces an output as a weighted combination of those states.The node dynamics use an interconnection matrix, input mask, feedback gain, and input gain; output weights are trained to minimize mismatch with the target.
- Reservoir-computing architecture: The experimental system implements a single nonlinear node with delayed feedback, an architecture chosen for experimental simplicity because it requires few components.The implementation uses a simple adjacent-node coupling rather than a randomly generated connection matrix, provided the feedback dynamics remain stable.
- Delay-loop operation: Desynchronizing the input and loop round-trip time couples each internal variable to the adjacent one, and this second approach is used in the experiment.The input is held for T, divided into N intervals of duration θ, and sequentially drives the reservoir using masked values.
- Optical hardware: The reservoir uses incoherent light in a 1.6 km fiber loop containing an SOA and tunable attenuation, with 50 internal variables and SOA saturation as the nonlinear function.The loop round-trip time is 7.9437 µs, the input timescale is 7.7880 µs, and θ is 155.76 ns.
- Optical hardware: A broadband SLED, Mach–Zehnder modulator, attenuator, photodiode, and digitizer provide optical driving, signal recording, and post-processing at 200 MSamples per second.The recorded circulating intensity is sampled through a 10% fiber-coupler tap for measurement of the internal states.
- Optical nonlinearity and dynamics: The measured SOA response bends above its saturation level, while the reservoir response exhibits nonlinear amplification and fading memory.The SOA output power also increases with injection current, and the reservoir response differs for positive and negative input perturbations.
3. Operation mode of the all-optical reservoir
The operation mode adapts signed task signals to positive optical intensities, records the reservoir states, and trains a regularized linear readout offline. Testing then uses fixed readout weights on the remaining recorded trace.
- Input encoding: Masked input values are renormalized to [0,1] and converted through an arcsine-driven MZM voltage spanning [0,Vπ].This uses the MZM’s full dynamical range while removing nonlinearity from its transfer function.
- State measurement: The recorded intensity trace is rescaled to [-1,+1], and each internal state is measured by averaging around its time-window midpoint over θ/2.This state measurement discards the leading and trailing edges of each interval, where abrupt jumps occur.
- Readout training: For each task, target outputs are defined from the input sequence, and a least mean square algorithm trains readout weights to minimize output mismatch.Ridge regularization is added to make training more robust against overfitting.
- Testing protocol: After training, fixed readout weights are tested on the remaining normalized intensity trace, while a warm-up sequence removes dependence on the initial steady state.The warm-up trace is excluded from training and serves to eliminate the reservoir’s initial memory.
4.Results
The study evaluates an all-optical reservoir using memory-capacity analysis, numerical modeling, and benchmark tasks. It achieves strong task performance with a simple system, while ASE noise limits some capacities and degrades performance relative to the optoelectronic reservoir.
- Operating point: The reservoir’s operating point is tuned through injection current, feedback gain, and input gain, with most tasks using 187mA injection current.The injection current shapes both the SOA nonlinearity and its ASE noise figure; feedback and input gains tune the reservoir dynamics.
- Modeling: The numerical model combines measured component transfer functions, SOA nonlinearity, and Gaussian internal noise to test whether the architecture imposes fundamental limitations.The model omits electronics bandpass effects and was mainly used to assess architectural and nonlinear limitations.
- Memory capacities: Memory capacities are lower than in the optoelectronic architecture, especially for cross memory and total memory approaching the 50-node maximum.The authors attribute this limitation to ASE noise from the SOA, which decreases the signal-to-noise ratio.
- Channel equalization: 12 to 16dB signal-to-noise ratio yields channel-equalization performance comparable to the optoelectronic reservoir, while higher SNR causes slight degradation.Even at higher SNR, measured SER remains one order of magnitude better than the cited bilinear filtering technique, and experiment agrees with simulation.
- Radar task: One- to ten-step radar prediction achieves performance similar to a previously reported 80-node reservoir for low sea state.The experiment uses a 50-node reservoir, evaluates low and high sea states, and measures prediction quality with NMSE.
- Spoken digit recognition: The best isolated spoken-digit recognition result is 3% WER, compared with 0.4% for the optoelectronic reservoir and 0 WER in discrete-time simulation.The experiment–simulation difference is attributed to ASE noise excluded from the simulations.
- Overall result: The work presents the first all-optical Reservoir Computer using a delayed feedback loop, an optical-amplifier nonlinearity, and standard off-the-shelf optical components.The authors report state-of-the-art capacities for several benchmark tasks despite the system’s simplicity.