Design and Optimisations of Metal-Oxide Artificial Synaptic Device Based Machine Learning Model

Synaptic device-based neural network models are increasingly favored for their energy-efficient computing capabilities. However, as the demand for scalable and resource-efficient computing solutions continues to grow, there is a pressing need to explore novel computational paradigms inspired by the human brain. Motivated by the ongoing imperative to enhance the accuracy performance of hardware-based neural network models to compete with software-based counterparts, this paper investigates the potential of memristor-based nanodevice, particularly
$TiO_{2}$
synaptic device, as a promising candidate for hardware-based neural network models and seeks to improve accuracy performance. The innovation of this work lies in the comparative analysis of optimization methods to improve the classification accuracy of hardware-based neural network models using
$TiO_{2}$
synaptic device. By investigating various optimization functions, including SGD, Momentum, RMSProp, Adam, and Adagrad learning methods, this study aims to provide insights into the effectiveness of these methods in enhancing the accuracy of
$TiO_{2}$
synaptic device-based neural network models. Experimental results demonstrate that the choice of optimization method significantly impacts the accuracy of the models, with the Adam algorithm achieving the highest classification accuracy of 92.39% on the MNIST dataset, showcasing the potential of optimized hardware-based models to advance machine learning applications, particularly in image processing and character recognition.