Dynamic Convolutional Neural Networks Based on Adaptive 2D Memristors

Abstract

Convolutional Neural Networks (CNNs) are pivotal in modern digital computing, particularly for tasks like image classification, inspired by the receptive fields of the human brain. Nevertheless, CNNs implemented on conventional digital computers face significant limitations due to inflexible kernels that cannot adjust to dynamic inputs, and the von Neumann architecture, which leads to inefficient data transfer between memory and processing units. This research presents a hardware-software co-designed solution, a Dynamic Convolutional Neural Network (dCNN), empowered by three-terminal adaptive two-dimensional (2D) memristors. These memristors consist of a vertical heterostructure integrating silver, an atomically thin insulator (CrPS₄), and graphene as a semimetal. This configuration allows for the dynamic tuning of conductive filament properties, emulating the heterosynaptic plasticity observed in biological neural systems. The three-terminal memristor design permits the dCNN to actively adjust kernel weights in its attention layer according to the input stimuli. The empirical tests demonstrate that image classification accuracy using our adaptive 2D memristor-enhanced dVGG reaches up to 94% on the CIFAR-10 dataset, which exceeds the performance of static VGG. Furthermore, the energy efficiency of our dVGG significantly outperforms that of GPUs, aligning more closely with the energy dynamics of the human brain in terms of both consumption and classification accuracy.