Spin wave interference-based efficient neuromorphic computing

Abstract

We demonstrate the design of a neuromorphic hardware, spin wave interference device (SWID), utilizing micromagnetic simulations, for performing feature extraction and classification of binary digit patterns. The SWID aims to reduce the weight computations in artificial neural network (ANN) implementations allowing for low power computing and faster inference. We achieve the direct classification of multibit binary input pulse schemes through synaptic behavior and interference of spin waves. We showcase the versatility of SWID's information processing capabilities across two-bit ranges, 4-bit and 6-bit binary digit data, by effectively controlling the nonlinearity and interference of spin waves with external input current pulses. The performance of the SWID with 4-bit and 6-bit digit pattern classification ability is tested for image recognition tasks with the Modified National Institute of Standards and Technology handwritten image database in a feed forward neural network. Though achieving 84.7% accuracy in image recognition, this SWID-based network reduces the weight computation by 99.4% as compared to the software-ANN, showcasing its capability for faster decision making. This huge reduction in computations offers great benefits to ANN applications in edge devices and memory constraint devices. These results underscore the potential of spin wave-based SWID in designing power efficient neuromorphic hardware.