Few-Shot Neuromorphic Vision in a Nonlinear Photonic Network Laser

With the growing prevalence of AI, demand increases for hardware that mimics the brain’s ability to extract structure from limited data. In the retina, ganglion cells detect features from sparse inputs via lateral inhibition, where neurons antagonistically suppress activity of neighbouring cells. Biological neurons exhibit diverse heterogeneous nonlinear responses, linked to robust learning and strong performance in low-data regimes. Here, we introduce a retinally-inspired photonic computing system where spatially-competing lasing modes in a random network laser act as heterogeneous, inhibitively-coupled neurons - enabling feature detection, few-shot classification, and segmentation. This silicon-compatible scheme harnesses heterogeneous excitatory and inhibitory nonlinear physical dynamics which give rise to emergent photonic computing behaviour, including parallel feature detection and strong performance when training data is scarce. We report 98.05% and 87.85% accuracy on MNIST and Fashion-MNIST, and 90.12% on BreaKHis cancer diagnosis - outperforming software CNNs including EfficientNetV2 and the vision transformer ViT in few-shot and class-imbalanced regimes with training sets of up to several hundred images. We demonstrate combined segmentation and classification on the HAM10k skin lesion dataset, achieving DICE and Jaccard scores of 84.49% and 74.80%. These results demonstrate the potential of random lasing networks as nonlinear photonic learning systems, and highlight the ability of heterogeneous nonlinear dynamics to support strong learning in challenging low-data scenarios.

💡 Research Summary

The authors present a neuromorphic vision platform that directly emulates retinal lateral inhibition using a random photonic network laser. A 150 µm InP‑on‑Si random laser, fabricated by wafer‑bonding and electron‑beam lithography, consists of a Voronoi‑type waveguide graph where hundreds of spatially overlapping lasing modes coexist. Input images are patterned onto the device with a 633 nm, 200 fs pump laser via a digital micromirror device (DMD). Each 4×4‑pixel window of the image selectively excites gain in a small region of the network; the resulting gain distribution determines which modes reach threshold and how strongly they lase. Because many modes share the same waveguide paths, they compete for the finite gain—a nonlinear “hole‑burning” effect that simultaneously provides excitatory (mode amplification) and inhibitory (mode suppression) dynamics. This dual nonlinearity mirrors the excitatory‑inhibitory coupling of retinal ganglion cells and yields a set of spectrally distinct output channels that act as parallel feature maps. In experiments, ten distinct modes each respond preferentially to different spatial features (horizontal edge, vertical edge, corners, etc.), producing ten simultaneous feature maps from a single raster scan. Simulations with the netSALT model predict over two hundred possible modes, confirming the high‑dimensional latent space inherent to the random laser.

For classification, the authors treat the intensity of each mode (or a subset) as a feature vector and feed it to simple linear classifiers or shallow multilayer perceptrons. Remarkably, with only a few hundred training samples, the photonic system achieves 98.05 % accuracy on MNIST, 87.85 % on Fashion‑MNIST, and 90.12 % on the BreaKHis breast‑cancer dataset—outperforming large‑scale software models such as EfficientNet‑V2‑B0 (7.9 M parameters) and ViT‑B/16 (86 M parameters) under the same low‑data regime. Moreover, the same hardware is used for combined classification and segmentation on the HAM10k skin‑lesion dataset, delivering a Dice score of 84.49 % and a Jaccard index of 74.80 %.

The key contributions are: (1) embedding both excitatory and inhibitory nonlinearities in a physical substrate, enabling energy‑efficient, parallel feature extraction; (2) leveraging the intrinsic heterogeneity of random lasing modes to create a high‑dimensional, expressive feature space that supports few‑shot learning; (3) demonstrating that a compact on‑chip photonic device can match or exceed state‑of‑the‑art deep‑learning models on challenging biomedical tasks with severely limited training data. Limitations include reliance on off‑chip pump lasers and spectrally resolved detectors, and a mode count constrained by the physical size and disorder of the network. Future work should integrate on‑chip pumping, on‑chip spectrometers, and scalable fabrication to realize fully integrated photonic neuromorphic processors for edge AI applications.