SpikeMT: A Spike-Driven Memristive Transformer Circuit for Heterogeneous Data Processing

Given the extensive applications of deep neural networks in various scenarios, research efforts have increasingly concentrated on deploying neural networks on edge devices to mitigate data transmission costs. However, a significant disparity exists between the hardware capabilities of edge devices and the computational power requirements of deep neural networks. In this work, we propose a memristive neuromorphic network with a reconfigurable spike transformer circuit (SpikeMT). It offers a low-power intelligent solution for edge devices. Additionally, it facilitates the deployment of multipurpose neuromorphic networks. A reconfigurable memristive spiking neuron is specifically designed for spiking transformers, enhancing the biological characteristics of artificial neuronal circuits. Furthermore, the memristive neurons serve as the core operators in the proposed memristive spiking self-attention circuits. They facilitate the integration of the self-attention mechanism within a spike-driven paradigm without the need of analog-to-digital converters. Notably, heterogeneous data can be processed at the edge without altering the hardware architecture of SpikeMT. Experimental results demonstrate that the proposed SpikeMT is advantageous in computational efficiency and supports vision tasks, promoting the development of edge AI devices.