Adaptive Pulse-Controlled Learning in Differential 1T1R Memristor Crossbar for Neuromorphic Computing

Abstract

Neuromorphic computing has been proposed as an exciting paradigm to accomplish energy-efficient artificial
intelligence by replicating the working principles of biological neural networks. Here, the memristor has attracted a lot of interest
because of its capability of mimicking the synaptic behaviour by means of programmable conductance states. Nonetheless,
traditional neural architectures using memristors have a few drawbacks such as low precision of weights, device variability and
inefficient programming schemes that limit their direct application to large systems. The current work suggests a pulse-controlled
learning scheme based on an adaptive pulse that is embedded into a differential 1T1R memristor crossbar design to improve the
accuracy of programming synaptic weights and energy efficiency. The overall aim is to actively control the conductance of
memristors with feedback-based pulse width modulation, which allows neural network weights to be mapped accurately,
nonlinearity and drift effects to be reduced. The proposed design is a hybrid analog design that includes a parallel vector-matrix
multiplication with a differential memristor crossbar, a pulse generation unit with adaptive programming and an analog front-end
that includes a differential amplifier and comparator to process and activate signals. There is also a write-verify scheme which
iteratively varies pulse width according to real-time measurements of conductance, which ensures that the weight values
converge to desired values. The conductance update dynamics are modeled mathematically in order to optimize pulse parameters
in order to enhance performance of learning. Simulation findings indicate that the proposed adaptive pulse-controlled strategy
attains a higher weight programming precision, lower power usage, and a shorter convergence time than fixed-pulse technique.
Moreover, the architecture has a better tolerance to device variability, and can be used in scaleable implementations of
neuromorphic hardware. These results indicate that adaptive pulse modulation is effective in supporting the future of memristor
based neural network designs in neural computing applications at the edge.