Image classification using optical spiking neural networks based on VCSEL with saturable absorption region neurons

Neuromorphic computing systems, inspired by the human brain's efficient neural synchronization, show significant potential for energy-efficient image classification tasks. We propose a neuromorphic optical framework utilizing vertical cavity surface-emitting lasers with saturable absorber regions (VCSEL-SA) for image classification tasks. A physics-based model simulates optical spiking dynamics in VCSEL-SA neurons. The convolutional neural network (CNN) is designed based on the ResNet-20 architecture, in which VCSEL-SA optical spiking units are implemented to function analogously to conventional leaky integrate-and-fire (LIF) neurons. The spiking dynamics are encoded through temporal correlation of laser spikes, where input images are converted into spike trains via latency coding-pixel intensity is inversely mapped to spike emission latency. In the proposed architecture, each residual block integrates a VCSEL-SA layer that replaces traditional ReLU activations, with synaptic weights trained via surrogate gradient descent. The proposed optical spiking neural network (OSNN) processes and evaluates diverse source images through spike-based feature encoding, followed by hierarchical classification via VCSEL-SA neuron layers under calibrated photonic dynamics. Numerical simulations demonstrate that the proposed system achieves a classification accuracy of 92.9% on the CIFAR-10 dataset. This result validates the feasibility of VCSEL-SA OSNN for high-precision visual processing.