Compact classification using the biomimetic properties of ultrafast spiking microlaser neurons

Machine learning using spiking neurons offers high computational efficiency with minimal resources, enabling sparse coding and brain-inspired methods. Photonic systems potentially enable very fast classifications but, despite recent progress, demonstrations are still challenging to implement. While biomimetic properties of optically spiking neurons have been shown theoretically and experimentally, exploring how these characteristics can be used for a classification task has rarely been considered. Simulations of such architectures are hindered by the complexity of the numerical modeling necessary to accurately describe their properties. Here we show that a surrogate model based on a machine learning algorithm can accurately and efficiently describe a microlaser neuron’s biomimetic behavior for a specific targeted application. This enables the training of a compact, energy efficient spiking photonic architecture that takes advantage of biomimetic properties to detect important features of the data. The resulting classification accuracy is comparable to that of a multi-layer perceptron with higher structural complexity. We anticipate that our method is applicable to any spike-based excitable neuron hardware that exhibits similar biomimetic properties, enabling simulations of more complex biomimetic neuron architectures and ultimately more efficient hardware.