Source-linked AI summary
Large-Scale Optical Reservoir Computing for Spatiotemporal Chaotic Systems Prediction
Mushegh Rafayelyan, Jonathan Dong, Yongqi Tan, Florent Krzakala, Sylvain Gigan
TL;DR
Large reservoirs are difficult to realize conventionally because computation and memory scale quadratically with network size. This paper develops an optical reservoir using parallel photonic processing and tests it on large Kuramoto-Sivashinsky chaotic datasets, finding linear overall scaling from electronic overheads while optical computation remains independent of reservoir size.
Problem
Conventional computers face quadratic computation-time and memory growth when realizing large reservoir networks.
Method
The paper uses an optical reservoir with SLM phase encoding, scattering-medium mixing, camera readout, and recursive prediction on Kuramoto-Sivashinsky datasets.
Results
The optical network scales to a million nodes, with computation time growing linearly from electronic overheads while optical matrix multiplication remains independent of reservoir size.
Takeaways & Limitations
The approach opens the prospect of high-speed, low-energy prediction on very large datasets such as turbulence.
Abstract
from arXiv · showhide
Reservoir computing is a relatively recent computational paradigm that originates from a recurrent neural network and is known for its wide range of implementations using different physical technologies. Large reservoirs are very hard to obtain in conventional computers, as both the computation complexity and memory usage grow quadratically. We propose an optical scheme performing reservoir computing over very large networks potentially being able to host several millions of fully connected photonic nodes thanks to its intrinsic properties of parallelism and scalability. Our experimental studies confirm that, in contrast to conventional computers, the computation time of our optical scheme is only linearly dependent on the number of photonic nodes of the network, which is due to electronic overheads, while the optical part of computation remains fully parallel and independent of the reservoir size. To demonstrate the scalability of our optical scheme, we perform for the first time predictions on large spatiotemporal chaotic datasets obtained from the Kuramoto-Sivashinsky equation using optical reservoirs with up to 50 000 optical nodes. Our results are extremely challenging for conventional von Neumann machines, and they significantly advance the state of the art of unconventional reservoir computing approaches, in general.
I. INTRODUCTION
Reservoir computing offers a route to large neural networks, but conventional realization is challenging because network-scale computation is difficult. Optical approaches exploit parallelism and scalable photonic coupling, and this work extends them to large Kuramoto-Sivashinsky predictions while measuring their scaling behavior.
- Reservoir computing simplifies recurrent-network training by fixing random input and internal connectivity matrices.
- Optics is promising for large neural networks because it provides parallel processing, light-speed data handling, and low energy consumption.
- Spatially distributed photonic reservoirs process large-scale inputs without sacrificing computation speed, while scattering media provide random coupling and parallel processing.
- The work experimentally applies optical reservoir computing to multidimensional Kuramoto-Sivashinsky chaotic predictions and evaluates computation-time scaling against conventional computers.
- The optical scheme shows nearly size-independent optical computation, with mild linear scaling attributed to electronic overheads, whereas conventional computation scales quadratically.
II. CONVENTIONAL RESERVOIR COMPUTING
Conventional reservoir computing injects inputs into a high-dimensional dynamical reservoir whose fixed random connections generate evolving states. The reservoir’s memory depends on its node count and nonlinear activation.
- An input vector of dimension D_in is injected into a reservoir state vector of dimension D_res, the number of reservoir nodes.
- The input and internal connection matrices are randomly initialized and fixed throughout the reservoir-computing process.
- Each reservoir state recursively evolves through an elementwise nonlinear function applied at discrete input time steps.
- Each subsequent reservoir state retains exponentially decaying information about previous states and injected inputs.
- Reservoir memory capacity is mainly determined by the number of reservoir nodes and the nonlinear activation function.
RMSE ¼
Reservoir computing trains output weights to map reservoir states to desired outputs and then uses those outputs for recursive future prediction.
- The output is computed as o(t) = W_out · r(t), combining reservoir states through trained output weights.
- Linear regression adjusts W_out so the computed output approaches the desired output, with regularization available to reduce overfitting.
- Only output weights are modified during training; random input and reservoir weights remain fixed and project inputs into a high-dimensional space.
- For prediction, training targets the next input time step, after which each prediction replaces the subsequent input in a feedback loop.
III. OPTICAL RESERVOIR COMPUTING
The optical reservoir encodes inputs and reservoir states on a phase-only SLM, mixes them through a scattering medium, and reads out the next state with a camera. Its recursive optical model extends conventional reservoir computing with phase encoding and complex-valued random matrices.
- A phase-only SLM encodes the input and reservoir state together into the spatial phase profile of light.
- Input and reservoir states are encoded with tunable gray-level ranges and macropixel sizes on the SLM.
- The scattering medium performs dense random linear mixing, while the camera provides a nonlinear readout of the next reservoir state.
- The optical recursion uses a function F that maps encoded input and reservoir-state matrices to the camera-detected next reservoir state.
- The optical model includes camera-intensity and SLM-phase nonlinearities, with complex-valued W_res and W_in matrices.
- Training remains a conventional-computer linear regression, while prediction feeds each one-step output back through the SLM as the next input.
- The optical reservoir is applied to Kuramoto-Sivashinsky spatiotemporal chaotic-system prediction.
IV. EXPERIMENTAL RESULTS
Optical reservoir computing predicts Kuramoto-Sivashinsky spatiotemporal chaos while scaling to larger reservoirs than conventional implementations. Prediction quality depends on matching reservoir size to system size, and the optical scheme gains speed and memory advantages despite electronic bottlenecks and stability limits.
- KS prediction setup: For L = 22, the experiment uses Nx = 64 spatial points, Nt = 90500 time steps, and an optical reservoir with Dres = 10^4 photonic nodes.The first 9 × 10^5 time steps are used for training.
- Optical reservoir configuration: The optical reservoir is tuned with s_res = s_in = 0.5, p_res = 64, p_in = 10 000, and λ = 0.07 for regularized linear regression.These settings encode input and reservoir states over 0 to 128 and maintain equal importance ratios on the SLM.
- Prediction evaluation: NRMSE is evaluated over 100 training and testing KS dataset pairs, with a mean curve obtained by averaging the individual test-sample errors.Prediction performance varies significantly across test samples.
- Scaling with system size: Prediction performance decreases rapidly as the KS spatial size L increases when the reservoir remains fixed at Dres = 10^4.The tested system sizes are L = 12, 22, 36, 60, and 100.
- Comparison with conventional computing: Conventional reservoir computation time and memory grow quadratically with node count, whereas the optical scheme can exploit all-to-all random mixing without large operating memory.Strongly scattering media are proposed as a route toward reservoirs with millions of photonic nodes, although measurement time remains a limitation.
- Scaling with reservoir size: For L = 60, increasing the reservoir from Dres = 10^4 to Dres = 5 × 10^4 recovers prediction performance, especially in intermediate temporal regions.These regions correspond to predictions that are neither too simple nor too difficult.
- Computation scalability: At Dres = 50 000, the optical implementation is about 4 times faster than the GPU and 100 times faster than the CPU, near their memory limits.The optical computation remains parallel, while camera-to-SLM electronic feedback produces linear time growth with reservoir size.
V. DISCUSSION AND CONCLUSION
The optical reservoir computing network scales to large chaotic prediction tasks while keeping optical matrix multiplication parallel and reservoir-size independent. Experiments identify electronic feedback as the source of linear computation-time growth and demonstrate million-node scalability.
- Discussion and conclusion: 10^6 photonic nodes can be hosted in the proposed optical reservoir network, with an estimated 10^14 operations per second.The estimate assumes about 10^12 operations per iteration and 100 Hz SLM and camera speeds.
- Discussion and conclusion: Electronic feedback causes computation time to grow linearly with reservoir size, although optical information processing remains fully parallel.
- Discussion and conclusion: The optical network performs first-of-its-kind predictions on large multidimensional chaotic datasets using the Kuramoto-Sivashinsky equation.
- Discussion and conclusion: The optical matrix multiplication is independent of reservoir size and requires no memory storage, supporting scaling to very large networks.
- Discussion and conclusion: The authors report that comparable prediction performance requires increasing the optical network size for larger chaotic systems.
VI. CODE AND DATA AVAILABILITY
The paper states that all data and codes are available through Ref..
- Code and data availability: All data and codes can be accessed at Ref..
APPENDIX A: EXPERIMENTAL SETUP
The setup uses a 532 nm laser, polarization optics, and a liquid crystal SLM to encode the input and reservoir state into the reflected beam’s phase profile.
- Experimental setup: A 532 nm laser beam is expanded tenfold before entering the polarization and spatial-light-modulation stages.
- Experimental setup: A polarizer and half-wave plate align the light with the SLM’s extraordinary axis for pure phase shaping.
- Experimental setup: The SLM receives the concatenated input i(t) and reservoir state r(t), encoding them into the reflected beam’s phase spatial profile.
APPENDIX B: LIGHT SCATTERING
Light scattering through a thick disordered medium produces a random speckle pattern and implements a dense transmission matrix between inputs and detectors. This matrix can be used for high-dimensional reservoir computing without explicitly measuring or storing it.
- Light scattering: Scattering through a thick medium creates a speckle pattern from interference among many complex scattering paths.
- Light scattering: The transmission matrix maps incident electric fields to detector outputs and is characteristic of the complete optical setup.
- Light scattering: A thick disordered medium between the SLM and camera produces a dense random transmission matrix that can be measured experimentally.
- Light scattering: The optical system can exploit the transmission matrix’s large dimensionality without measuring its explicit form by using reservoir computing algorithms.
APPENDIX C: KURAMOTO-SIVASHINSKY TIME SERIES
Appendix C describes the Kuramoto–Sivashinsky system used to generate chaotic time series for reservoir input and relates its dynamics to spatial-domain size through the largest Lyapunov exponent.
- The Kuramoto–Sivashinsky equation models nonlinear chaotic systems with intrinsic instabilities, including laminar flame-front instabilities and hydrodynamic turbulence.
- The scalar field is periodic over a spatial domain of size L, with solutions defined on [0, L).
- The equation is integrated on 64 equally spaced spatial points with a time step of Δt = 0.25.
- The resulting 64 time series are collected into the vector u(t) and used as the reservoir input.
- The largest Lyapunov exponent quantifies exponential divergence of nearby trajectories, while the spatial-domain size L strongly affects the system dynamics.
APPENDIX D: NONLINEAR ACTIVATION FUNCTION IN THE OPTICAL RESERVOIR COMPUTING SCHEME
Appendix D examines how adding a tunable nonlinear activation after the optical reservoir’s intensity readout affects prediction of Kuramoto–Sivashinsky datasets.
- The optical reservoir’s activation can be modified through camera saturation, gain, exposure time, or a camera look-up table.These approaches implement additional analog nonlinear transformations of the output intensity image.
- The standard intensity readout f(q) = |q|^2 is compared with the tuned nonlinear function f(q) = 1 − tanh(2.06|q|^2).
- Approximately 50% improvement in prediction performance is obtained with the appropriately tuned hyperbolic tangent nonlinearity.
- Figure 7 separates actual datasets, reservoir predictions, and errors for the standard and hyperbolic-tangent activation functions.