Oscillatory Neural Network with Tunable Frequency for Brain-Inspired Neuromorphic Computing

Abstract

We present a silicon transistor-based oscillator with frequency tunability (SOFT) for brain-inspired neuromorphic computing. It consists of a single transistor-based oscillator (1T-O) and a single transistor-based resistor (1T-R), which is a nonvolatile memory device that provides conductance modulation to enable frequency tunability. Because the 1T-O and 1T-R are homologous metal-oxide-semiconductor field-effect transistors (MOSFETs), they are structurally identical but function differently. We simultaneously integrated them on the same wafer using complementary metal-oxide-semiconductor (CMOS) fabrication due to their homogeneity. We then demonstrate template matching by resistively coupling two SOFTs and classify temporal signals using first-harmonic injection locking (FHIL) with four SOFTs, leveraging their frequency tunability. This SOFT is well-suited for high-density oscillatory neural networks (ONNs) with low-cost CMOS fabrication, offering significant potential for energy-efficient and areal scalability.