Analyzing Bridge Synaptic Neuron Circuit and Mapping to Memristor Crossbar for Neuromorphic Application

Memristor crossbar is a suitable architecture to perform vector-matrix multiplication (VMM) and additionally offers benefits like in-memory computing (IMC), low power and high density. When a memristor crossbar is used for neuromorphic computing, the weights are stored in the form of resistance/conductance. However, the nonlinear switching behavior of memristors prevents their efficient use for neural network applications. The main challenge with nonlinear storage is the precise configuration of the memristor resistance/conductance values. Furthermore, due to the resistive storage technique, memristors do not allow negative network weight. A memristor-bridge-synapse ensures linear weight change and handles negative weights. However, it cannot be directly mapped to the crossbar. In this paper, we propose a crossbar architecture suitable for bridge-synaptic neurons. Implementing the bridge-synaptic network in the crossbar allows parallel weight updates and reduces overall training and inference latency. The effectiveness of the proposed bridge-crossbar mapping is validated through experiments on MNIST, Fashion-MNIST, and CIFAR-10 datasets, demonstrating improved accuracy over conventional single-crossbar implementations while maintaining robustness and scalability for both single-neuron and multi-neuron networks.