Memristor Differential Pair Ternary Weight Neural (TWN) Network Architecture

Matrix multiplication is a computationally intensive task, and existing neural network training approaches often require substantial energy and time due to frequent data movement between memory and processing units in traditional von Neumann architectures. This inefficiency has prompted growing interest in alternative computational paradigms. In this work, we present the design and implementation of a ternary-weight memristive architecture employing memristor differential pairs for image classification tasks. The proposed system utilizes weight columns composed exclusively of ternary values (-1,0, and 1), thereby significantly reducing programming complexity and time. Simulation results demonstrate that, relative to full-precision implementations, the proposed architecture achieves enhanced computational efficiency while maintaining a classification accuracy of 96%, underscoring its potential for future edge computing applications.