Total: 1
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 bio-inspired 'retinomorphic' photonic system where spatially-competing lasing modes in a network laser act as heterogeneous, inhibitively-coupled neurons - enabling few-shot classification and segmentation. This compact (150 micron) silicon-compatible scheme addresses key challenges in photonic computing: physical nonlinearity and spatial footprint. We report 98.05% and 87.85% accuracy on MNIST and Fashion-MNIST, and 90.12% on BreaKHis cancer diagnosis - outperforming software CNNs including EfficientNet in few-shot and class-imbalanced regimes. 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 establish a new class of nonlinear photonic hardware for versatile, data-efficient neuromorphic computing.