Memristive blinking neuron enabling dense and scalable photonically-linked neural network

The rapid growth of artificial intelligence (AI) has driven a surging demand for powerful architectures whose capabilities scale with network size and dense neuron connections. However, the large footprint of CMOS neurons and the constraints on electric routing are significantly hindering the scale-up of the number of artificial neurons and their synaptic connectivity. Inspired by the biological neurons’ intricate and dense network structure, we introduce a 3-dimensional (3D) connectivity. Here, we demonstrate a 3D neural network (NN) with atomic scale memristive blinking neuron (MBN) linked by photons free from electrical routing constraints. Our MBN features a minimal footprint of 170×240 nm ² and emits photon pulses upon integrating a critical number of incoming electrical spikes, eliminating the need for the bulky peripheral circuit readout and electrical wiring for transmitting the signals. Such MBN not only offers the functionality of a spiking neuron but also constitutes a CMOS-compatible planar light source for intra-neuron communication, overcoming the bottleneck of plenary CMOS technology. The functionality of the MBN is demonstrated in two different photonically-linked NN applications. First, we test the MBN on a four-class classification task within the Google Speech dataset and show 91.51% accuracy. We then show its operation in a high-density artificial neuron array with a pitch of 1 µm, pushing the boundaries of optical diffraction, realizing an unprecedented neuron density on par with the human brain. An MNIST classification task is performed with a 92.27% accuracy. Our new photonically-linked NN significantly increases the neuron density as well as scalability, showing great potential for realizing next-generation AI hardware.