Source-linked AI summary
High performance photonic reservoir computer based on a coherently driven passive cavity
Quentin Vinckier, François Duport, Anteo Smerieri, Kristof Vandoorne, Peter Bienstman, Marc Haelterman, Serge Massar
TL;DR
Reservoir computing needs compact ways to process time-dependent signals while retaining strong performance across difficult analog tasks. This paper implements a coherently driven passive fiber-cavity reservoir with masked inputs and photodiode quadratic readout, achieving error rates as low as or lower than prior experiments across tested tasks while using low power. The paper also identifies a simple analytical model and a scope boundary for tasks requiring nonlinearities beyond quadratic order.
Problem
Reservoir computing seeks simple, flexible processing of time-dependent signals, while optical implementations still differ in flexibility, signal type, noise, and energy use.
Method
The paper experimentally implements a coherently driven passive fiber-cavity reservoir that combines delay dynamics, masked desynchronized inputs, and quadratic photodiode readout.
Results
Across all tested tasks, the experimental reservoir has error rates as low as or lower than previous experiments studying the same tasks.
Takeaways & Limitations
The experiment and its analytical model provide a simple, flexible, low-energy photonic reservoir-computing approach for analog signal processing.
Takeaways & Limitations
Tasks requiring nonlinearities higher than quadratic may not be well served by this architecture.
Abstract
from arXiv · showhide
Reservoir computing is a recent bio-inspired approach for processing time-dependent signals. It has enabled a breakthrough in analog information processing, with several experiments, both electronic and optical, demonstrating state-of-the-art performances for hard tasks such as speech recognition, time series prediction and nonlinear channel equalization. A proof-of-principle experiment using a linear optical circuit on a photonic chip to process digital signals was recently reported. Here we present a photonic implementation of a reservoir computer based on a coherently driven passive fiber cavity processing analog signals. Our experiment has error rate as low or lower than previous experiments on a wide variety of tasks, and also has lower power consumption. Furthermore, the analytical model describing our experiment is also of interest, as it constitutes a very simple high performance reservoir computer algorithm. The present experiment, given its good performances, low energy consumption and conceptual simplicity, confirms the great potential of photonic reservoir computing for information processing applications ranging from artificial intelligence to telecommunications
1. Introduction
The paper presents a coherently driven passive fiber-cavity reservoir computer for analog signals, combining delay dynamics with a linear optical circuit and quadratic photodiode readout. It reports flexible operation, low-loss passive implementation, and error rates matching or improving on prior experiments across tested tasks.
- Reservoir computing maps time-dependent inputs into high-dimensional nonlinear states, then trains a linear readout by minimizing mean square error.The method is presented as simple to train and applicable to tasks including speech recognition, channel equalization, seizure detection, robot control, and time-series prediction.
- The proposed reservoir combines delay-dynamical processing with a linear optical circuit whose photodiode readout supplies a quadratic nonlinearity.This architecture is implemented experimentally using a coherently driven passive fiber cavity processing analog signals.
- Unlike the cited photonic-chip experiment, the proposed architecture can tune internal variables and connection strength for task-specific performance.It also processes analog signals, whereas the cited chip experiment processed digital signals because of high-speed-electronics limitations.
- The passive cavity requires no amplifier or active cavity element, reducing a major noise source and energy consumption relative to prior delay-line experiments.The cavity has very low intra-cavity losses, supporting passive operation.
- Across all tested tasks, the experimental reservoir achieves error rates as low as or lower than previous experiments studying the same tasks.The comparison also includes many prior simulation and digital-algorithm results.
2. Operation principle
The reservoir uses masked, desynchronized inputs and a passive cavity recurrence to generate complex fading-memory states, which photodiodes square before linear readout. Training optimizes readout weights by regularized mean-square-error minimization, while performance depends on the tasks’ required nonlinear order.
- Input encoding: The input is held for duration T’, multiplied by a periodic mask, and desynchronized from the cavity by choosing T’≠T.A bias A0 is essential for some tasks, while avoiding common divisors between k and N produces richer unsynchronized dynamics.
- State dynamics: The cavity states are discretized as x_i(n)=A(t(n,i)) and updated from delayed states, masked inputs, bias, feedback gain, transmission, and phase detuning.The discrete recurrence uses separate expressions for k≤i<N and 0≤i<k.
- Readout: Photodiodes transform each complex cavity state into |x_i(n)|², and the reservoir output is a linear combination of these squared magnitudes.The photodiode therefore supplies the quadratic transformation used by the readout.
- Training and testing: During training, ridge-regularized readout weights minimize mean square error; during testing, fixed weights generate outputs for comparison with targets.The numerical model used later is given by the operation equations.
- Performance factors: When Δφ≠0, complex variables provide 2N internal degrees of freedom and richer dynamics, while the passive cavity’s weak noise supports high memory capacity.The absence of active elements is identified as the source of the weak noise.
- Scope: The architecture represents outputs as constant, linear, and quadratic fading-memory functions of prior inputs, so tasks requiring higher-than-quadratic nonlinearities may perform poorly.The bias A0 enables linear terms in the squared states.
3. Experimental implementation
The experiment uses a 1550 nm laser, Mach–Zehnder amplitude modulation, and a roughly 230 m passive fiber cavity with controlled feedback and phase detuning. Readout power can be reduced substantially without affecting performance, within the photodiode signal-to-noise constraint.
- Optical source: The optical input is generated by a continuous-wave 1550 nm laser whose coherence time exceeds the inverse cavity linewidth.Laser power is adjusted between 11 and 41 mW, with quoted power measured at the cavity entrance.
- Input modulation: A push-pull lithium-niobate Mach–Zehnder interferometer driven by an arbitrary waveform generator encodes the input in laser amplitude.The experiment uses either precompensated modulation or the modulator’s sinusoidal transfer function, with comparable reservoir performance.
- Bias control: Biasing is added through a DC electrode when required, while RF scaling controls the modulated input amplitude.Measured half-wave voltages are 7.54 V for DC and 7.27 V for RF.
- Cavity parameters: The approximately 230 m passive fiber cavity has T=1.13209 µs, uses k=1 and N=50 for most tasks, and refreshes outputs at approximately 0.9 MHz.The reported processing speed is limited by relatively slow electronics and the long fiber cavity.
- Feedback and stabilization: A 90/10 coupler injects the input, an attenuator tunes feedback gain over 0<α<0.806, and a counterpropagating control loop stabilizes cavity phase detuning.The control loop uses a photodiode, PID regulator, and piezo-electric fiber stretcher.
- Validation: Experimental results are compared with discrete-time simulations, with further setup and simulation details supplied in the appendix.
4. Results
The results section evaluates the reservoir experimentally and numerically on benchmark tasks, compares performance with published reservoir architectures, and averages repeated trials for most parameter settings.
- The evaluation uses benchmark tasks from the reservoir-computing community, generally with N=50 internal variables and a ridge parameter of 10^-4.The study scans several parameters, including feedback, phase detuning, bias, modulation, and regularization.
- Simulations guide experimental exploration toward parameter-space regions expected to provide good performance, and experimental results are compared with simulations there.
- For most tasks, each parameter choice is repeated ten times with ten different input series and the resulting performance is averaged.Speech recognition is excluded from this repetition procedure.
- The reported performances are compared with the best published results available for different reservoir-computer architectures.
- The paper provides brief descriptions of the evaluated tasks and refers readers elsewhere for more detailed task definitions.
A. Memory capacities evaluation
The reservoir is evaluated on linear, quadratic, cross, and total memory capacities, with performance generally comparable to or better than three experimental architectures. Its quadratic and cross memories are especially strong, while the experimental total capacity approaches the theoretical maximum.
- Memory-capacity measures: The task measures the reservoir’s ability to recall linear or nonlinear functions of previous inputs using linear, quadratic, and cross memory capacities.Total memory is the sum of these three capacities and theoretically cannot exceed the number of internal variables.
- Comparison with prior architectures: Quadratic memory is about 3 times larger than optoelectronic and SOA-based RC values and more than 5 times larger than the SA-based RC value.
- Comparison with prior architectures: Cross memory is slightly larger than the optoelectronic value, more than 6 times larger than the SOA-based value, and 2 times larger than the SA-based value.
- Comparison with prior architectures: The total memory capacity is comparable to the optoelectronic RC and larger than the SOA- and SA-based RCs.
- Overall performance: The experimental total capacity is very close to the maximum value of 50, while the noiseless simulation is extremely close to that maximum.The authors interpret the experimental result as indicating that little noise affects the experiment.
B. NARMA10
NARMA10 tests whether the reservoir can reproduce a nonlinear tenth-order system from random inputs. The reported error decreases with increased resolution and reservoir size, reaching strong experimental and algorithmic results.
- Task definition: NARMA10 is a benchmark task requiring reproduction of a nonlinear, tenth-order system driven by random inputs in [0,0.5].
- Evaluation procedure: The reservoir was trained for 1000 input steps and tested for the following 1000 steps, with NMSE used for performance measurement.Standard deviations were obtained by repeating the procedure 10 times.
- Experimental results: Using N=50 internal variables, experimental NMSE was 0.107±0.012 and simulation NMSE was 0.104±0.02.These results used a precompensated masked input signal and no bias.
- Resolution dependence: Increasing acquisition resolution from 8 to 14 bits yielded simulated NMSE=0.062±0.008.The authors attribute this improvement to reduced quantization noise in the simulation model.
- Dependence on reservoir size: With N=300, experimental NMSE was 0.0484±0.0095 and simulation NMSE was 0.0463±0.0142.
- Algorithmic result: The simple algorithm reached NMSE=0.0106±0.0030 with N=400, compared with reported simulation values of 0.022 at N=400 and 0.018 at N=520.
C. Nonlinear channel equalization
The nonlinear channel equalization task recovers symbols from a noisy nonlinear multipath channel. Under the reported test conditions, the passive coherent reservoir outperforms previous architectures and achieves zero symbol errors at high SNR.
- Task definition: The task recovers randomly selected symbols from a standardized nonlinear multipath RF channel corrupted by Gaussian noise.Symbols take values in {-3,-1,1,3}, with SNR ranging from 12 to 32 dB.
- Evaluation metric: Performance is evaluated using Symbol Error Rate, defined as the fraction of misclassified symbols.
- Evaluation procedure: The experiment used 50 internal variables, 3000 training samples, and 6000 test samples, with standard deviations obtained from 10 repetitions.
- Results: At 28 dB and 32 dB SNR, both simulation and experiment achieved SER of 0%, correctly identifying all 60000 symbols.
- Results: The passive coherent RC outperforms previous architectures under the same test conditions.
D. Isolated spoken digits recognition
The reservoir computer accurately recognizes isolated spoken digits, including under severe babble noise, with experimental results closely matching simulations. Its performance is competitive with or better than prior reservoir-computing experiments and algorithms on comparable settings.
- Evaluation: The benchmark used ten digits pronounced ten times by five female speakers, with a second condition adding babble noise to reach 3dB SNR.The speech data came from the NIST TI-46 Corpus and was pre-processed using the Lyon cochlear ear model.
- Evaluation: Performance was evaluated using five-fold cross validation over five randomly chosen subsets of 100 words.Word Error Rate (WER) denotes the fraction of misclassified digits.
- Noisy recognition: 0.8(±0.8)% WER was obtained experimentally at 3dB SNR with N=500 under babble noise.No prior experimental result was available for direct comparison; simulation produced WER=0.6(±0.9)% with N=500.
- Algorithmic performance: WER=0 was reached by the corresponding simple algorithm with N=90 for noiseless signals and N=350 for noisy signals.These results used precompensation, no bias, and scans over α, Δφ, and the ridge parameter.
5. Conclusions
The study demonstrates a coherent-light passive fiber reservoir computer for analog processing, combining tunability, low noise, and strong benchmark performance. Its conceptual simplicity also yields a compact algorithm, while stabilization remains the main practical challenge.
- 5. Conclusions: The experiment implements a passive linear fiber reservoir computer using coherent light for analog signal processing.Its key parameters are tunable, enabling task-specific operating points.
- 5. Conclusions: Stabilization was the main challenge because the reservoir is a long interferometer containing an approximately 230m fiber cavity.Faster electronics may enable smaller, potentially integrated cavities and simpler stabilization.
- 5. Conclusions: Across tested tasks, the experiment achieved error rates as low as or lower than previous experiments.The conclusion highlights this pattern across the evaluated benchmarks.
- 5. Conclusions: No earlier experiment had exceeded a linear shift register on the NARMA10 task before this work.NARMA10 was described as a major challenge for prior experiments.
- 5. Conclusions: The discrete-time equations define a simple high-performance reservoir algorithm combining a simple interconnection matrix, a linear reservoir, and a nonlinear output layer.The authors note that comparison with other reservoir-computing algorithms remains necessary.
- 5. Conclusions: The architecture is presented as an important milestone for future progress in photonic reservoir computing.This conclusion is based on the experiment's reported performance and conceptual simplicity.
1. Numerical simulations
The setup uses a stabilized, polarization-tunable passive fiber cavity, with simulations incorporating experimentally relevant component limitations. Noise isolation and the long cavity support operation with relatively slow electronics.
- Simulation model: The simulation includes the 14-bit AWG, 8-bit oscilloscope, RF-amplifier saturation and bandwidth, and photodiode low-cutoff effects.These features are incorporated alongside the discrete-time reservoir equations and nonlinear preprocessing when relevant.
- Cavity stabilization: A ~230m cavity has roundtrip time T=1.13209µs and uses a piezoelectric stretcher to compensate thermal, vibrational, and phonic phase noise.The cavity's phase must be stabilized because of its length.
- Polarization and detuning: The polarization controller tunes the cavity Jones matrix and the phase offset between its two polarization eigenmodes.The reservoir-state and counterpropagating control signals are injected into different polarization eigenmodes, while a PID stabilizes the cavity on a control-resonance slope.
- Resonance measurement: Cavity resonances are measured by scanning the phase with a piezoelectric stretcher while recording output power from counterpropagating CW signals on the two polarization eigenmodes.The optical attenuator is set to maximum transparency during this measurement.
- Noise isolation: The cavity components are enclosed in nested, stone-wool-lined aluminum boxes and mounted on sorbothane sheets to reduce vibration and phase noise.The boxes target frequencies beyond the PID's 8.3kHz refresh capability; the full experiment also rests on an air-cushioned optical table.
- Implementation scope: The 230m fiber cavity was chosen to accommodate relatively slow AWG and photodiode electronics, while smaller cavities are intended for faster electronics.A smaller cavity is expected to be easier to stabilize and isolate from phase noise.