Improving the convergence of memristor-based neural networks

QUIÑONEZ, WALTER; MARÍA JOSÉ SÁNCHEZ; DIEGO RUBI

Virtual

Congreso; European Materials Research Society Meeting; 2022

Neuromorphic computing aims to emulate the architecture and information processing mechanisms of the brain. For this, new micro or nano-electronic devices (hardware) able to replicate the electrical behavior of synapses and neurons are needed. Memristors -metal/insulator/metal structures able to change their electrical resistance between different non-volatile states upon the application of electrical stress- are capable of electrically reproducing the adaptive –analog- synaptic weight of brain synapses [1]. It has been shown that memristor arrays with cross-bar architecture could be a possible physical implementation of neural networks [2]. In this work, starting from the experimental potentiation/depression curves measured on different metal/manganite systems, we analyze by means of numerical simulations how the physical constraints present in real memristive systems, such as limited conductivity windows or non-linear and discrete potentiation/depreciation curves, affect the convergence and accuracy of simple neural networks for image recognition. We also develop strategies to accelerate the convergence by introducing stochasticity in the actualization of the synaptic weights during the network training. The results obtained here are expected to contribute to the optimization of hardware neural networks based on memristor cross-bar arrays.[1] S. Yu, Neuro-Inspiring Computing Using Resistive Synaptic Devices (Springer International Publishing, Cham, 2017).[2] M. Prezioso et al., Nature 521, 61 (2015)