Achieving superior accuracy in photonic neural networks with physical multi-synapses

Analog neural networks, which mimic numerical computations through energy-efficient physical transformations in hardware architectures, typically achieve lower accuracies than digital neural networks. We explore the potential of photonic neural networks to outperform digital counterparts. Unlike traditional analog computing, our extreme learning machine (ELM)-based photonic neural network operates with physical synaptic connections without relying on mathematical descriptions. Noteworthy accuracy enhancements are achieved through photonic multi-synaptic connections, going beyond conventional notions of network depth or nonlinearity. Experimental results on MNIST, Fashion-MNIST, and CIFAR-10 datasets demonstrate classification accuracies up to 99.79%, 98.26%, and 90.29%, respectively, outperforming digital counterparts and most reported hardware architectures. This underscores the transformative impact of photonic neural networks, especially with the pivotal role of photonic multi-synapses, in advancing intelligent devices and signal processing.