LAUSR

Memristive neuromorphic computing systems, which combine memory and processing units, are expected to achieve high energy efficiency. However, practical implementations often face challenges due to the complexity of complementary metal-oxide-semiconductor (CMOS) neurons and integration issues related to memristive synapses. In this work, a physical spiking neural network (SNN) system is introduced, featuring two silicon-based one-transistor (1T) neurons and an Al/ZrO2/TiO2/AlOx/Al (AZTA) memristive synapse for on-chip learning. The 1T neuron exhibits a natural latch effect, generating spikes that serve as both an encoder and a decoder. Meanwhile, the AZTA memristor enables analog weight updates through its inherent long-term potentiation and depression plasticity. The combined physical kernel enables algorithm-compatible on-chip learning and inference without the need for complex peripheral circuits, resulting in low energy consumption, high scalability, and dense integration. The compact two-transistor-one-transistor-one-resistor (2T-1T1R) kernel, which learns in real-time using a modified spike-time-dependent plasticity rule, can potentially achieve 91.53% classification accuracy on the MNIST data set, and 75.28% on the challenging Fashion-MNIST data set, in an unsupervised manner if the device uniformity is sufficiently high, as demonstrated through Python simulations. It delivers competitive accuracy compared to the latest CMOS-based SNN processors while significantly reducing energy use (a 1784× decrease per update and a 1350× decrease per inference) with minimal hardware complexity.