The continuous development of artificial intelligence requires a more powerful computing architecture. However, the huge footprint and electrical conductance routing limitations of complementary metal oxide semiconductor-based systems hinder the expansion of traditional artificial neurons and their synaptic connections. We have demonstrated that memory-driven flashing neurons can be used to construct scalable photonic link three-dimensional neural networks. Our artificial neurons are based on silver/poly(methyl methacrylate)/silver metal-insulator-metal intramolecular switching junctions. Their resistance switches rely on atomic-scale filamentary dynamics and emit photon pulses by integrating a critical number of incident electrical spikes, thereby eliminating the need for large peripheral circuit readout and electrical wiring for signal transmission. Using memory-driven flashing neurons with a coverage of 170 nanometers by 240 nanometers, we constructed a three-dimensional spike neural network with photonic links. We demonstrated that this network can perform four-class classification tasks with an accuracy of 91.51% on the Google voice dataset. We also created a high-density artificial neuron array with a 1 micron spacing and proved that it can complete the MNIST classification task with an accuracy of 92.27%. This research was published in Nature Electronics under the title "Photonically linked three-dimensional neural networks based on memristive blinking neurons".
Reference News: 
DOI: 10.1038/s41928-025-01529-5