Three-Dimensional Memristor-Based Crossbar Architecture for Capsule Network Implementation

Abstract

Although the Capsule Network (CapsNet) has a better proven performance for the recognition of overlapping digits than Convolutional Neural Networks (CNNs), a large number of matrix-vector multiplications between lower-level and higher-level capsules impede efficient implementation of the CapsNet on conventional hardware platforms. Since three-dimensional (3-D) memristor crossbars provide a compact and parallel hardware implementation of neural networks, this paper provides an architecture design to accelerate convolutional and matrix operations of the CapsNet. By using 3-D memristor crossbars, the PrimaryCaps, DigitCaps, and convolutional layers of a CapsNet perform the matrix-vector multiplications in a highly parallel way. Simulations are conducted to recognize digits from the USPS database and to analyse the work efficiency of the proposed circuits. The proposed design provides a new approach to implement the CapsNet on memristor-based circuits.