Reservoir neuromorphic computing based on spin-orbit coupling in an organic crystal resonator

TL;DR

This work demonstrates reservoir neuromorphic computing realized in a spin-orbit-coupled photonic resonator built from a 2D BPDBNA organic crystal, enabling nonlinear separation of inputs via interference. By exploiting OSHE to expand the reservoir output dimensionality through polarization, the approach achieves a 10× reduction in network size and a 3× speedup for MNIST-style tasks, while maintaining accuracy. The low-power, compact, and chip-friendly design suggests a general pathway to improve photonic reservoir computing, with training times reduced by up to ~30× in the symbol tasks and sub-microsecond inference after training.

Abstract

Neuromorphic computing is at the basis of the recent progress in artificial intelligence. But the progress is accompanied with increasing demands in computational resources and power supply. Reservoir neuromorphic computing uses a non-linear physical system to replace a part of a large neural network. The advantages can include reduced power consumption and faster learning. We show that the interference in an organic crystal waveguide resonator leads to efficient separation of optical patterns, allowing a significant reduction of the size of the neural network and an acceleration of the learning process. For more complex symbols, extending the reservoir output dimension thanks to spin-orbit coupling, we achieve a 10-times reduction of the network size and a 3-fold speedup. Our work suggests a general path for the performance improvement of photonic reservoir computing systems.

Figures (13)

• Figure 1: Reservoir neuromorphic computing with a spin-orbit-coupled optical resonator. a) A multi-layer neural network. The training involves adjusting weights in all layers. b) A shallow neural network using a nonlinear physical system as a reservoir. The training only involves the output layer. c) Photonic spin-orbit coupling: the dispersions of the TE and TM modes in a planar resonator.(Experimental measurements in Figure S5) d) The direction of the dipole moment in BPDBNA crystals e) A scheme of a hexagonal optical resonator with the output signal collected from the sides.
• Figure 2: Efficiency of experimentally implemented reservoir neuromorphic computing. a,b) The input symbols on the surface of the resonator; c) The scheme of the system: the reservoir + a $3\times 10$ shallow network. One of the input symbols with the definition of the intensity integration sectors ($I_1$, $I_2$, $I_3$); d) The rapid increase of the prediction accuracy during training with reservoir. e) The separation and clustering properties: the output signal in the 3D space spanned by $I_1$, $I_2$, $I_3$ for each of the 10 symbols indicated by the colorbar. f) The confusion matrix plot demonstrating a very high final accuracy.
• Figure 3: Reservoir computing applied to the MNIST dataset. a) The total intensity distribution created by one of the MNIST symbols in the resonator, with the sectors indicating the intensity integration regions from the edges. b) One of the sectors of the sample, with the emission from the edge decomposed into 6 polarization channels. c) The evolution of accuracy of the trained $72\times 10$ network with the number of steps used in training (solid black line - power law fit). d) The accuracy (left axis, red) and the training time (right axis, blue) as functions of the number of channels in the reservoir. The case without reservoir (784 bit image) is shown on the right as a reference for comparison. Points with error bars (standard deviation) correspond to the experiment, solid lines - power law fits. e) The confusion matrix, corresponding to an overall accuracy of 95.6% for this particular network.
• Figure S1: The structure of the (2Z,2’Z)-3,3-([1,1’-biphenyl]-4,4’diyl)bis(2-(naphthalen-2-yl)acrylonitrile) (BPDBNA) molecule. The small balls represent carbon atoms (gray), hydrogen atoms (white), and nitrogen atoms (purple), respectively.
• Figure S2: Experimental setup allowing to obtain polarization-resolved complete state tomography. BS: beam splitter; L1-L4: lenses; M1: mirror. The red beam traces the optical path of the reflected light from the sample at a given angle.