Continual Learning Electrical Conduction in Resistive‐Switching‐Memory Materials

Abstract

The authors apply the concept of continual learning to modeling conductive‐filament growth in resistive‐switching materials (RSM). The approach permits computation of compliance current without knowing the geometries of conductive filaments and switching behaviors. This avoids the need to retrain the entire dataset when additional compliance currents are considered and is thus ideal for resistive switching (RS) thin films, doped layers, and other material systems. Computation of compliance current is consistent with experimental data for a wide range of parameters and learning tasks and demonstrates switching behavior not captured by traditional models. Lesion calculations elucidate the brain‐inspired‐modification‐facilitated increase in compliance‐current‐computation‐accuracy. A state‐of‐the‐art performance on challenging compliance‐current learning tasks without storing data is achieved (“brain inspired replay (BIR) – gating based on internal context (Gat)” method, ≈89%; above a benchmark of ≈50% for synaptic‐intelligence (SI)/elastic‐weight‐consolidation (EWC) methods), and it provides a novel model for computing compliance current based on replay of the brain. This intuitive approach combined with a simple solver tool, allows researchers with little computation experience to perform realistic and accurate modeling.

Nonvolatile Memristive Materials and Physical Modeling for In‐Memory and In‐Sensor Computing

“…[148] The resistive switching is a result of the redox reaction and ion migration, which are induced by electric potentials, chemical potentials, and temperature gradients over reaction coordinates, in the general theory of the resistive switching process. [149][150][151][152][153] When ions are electrically or thermally stimulated, anions including oxygen ions or the corresponding oxygen vacancies become more movable compared to cations for various dielectric materials, such as perovskites and transition-metal oxides. The variation in the metal-cation valence in the dielectric material alters electrical conductivities when oxygen anions move toward the anode upon the application of an electric field (Figure 4a).…”

Toward Memristive Phase‐Change Neural Network with High‐Quality Ultra‐Effective Highly‐Self‐Adjustable Online Learning

“…Currently, the focus for enhancing the speed and power performance of artificial intelligence tasks is on training, driven by DOI: 10.1002/apxr.202300085 the utilization of hardware accelerators for achieving machine-learning inference in deep neural networks (DNNs) . [1][2][3][4][5][6] Theoretical calculations have revealed that in-memory analogue computing using memristive crosspoint architecture result in a substantial energy, area and time benefit compared to that using a conventional digital complementary metal-oxide semiconductor (CMOS) system. [7][8][9][10][11][12] However, traditional memristive experimental studies have been constrained to reduced problem types.…”

“…Memristive hardware exhibits a small energy per conductance update, rapid switching time and excellent device miniaturization. [13][14][15][16][17][18] Computing and data storage have been achieved using memristive hardware through alterable conductance levels, which enables memory and processing to be combined in a parallel-based design. [19][20][21][22][23][24] Experiments have demonstrated the maximum number of distinct conductance levels utilized for neural-network (NN) learning for different memristive elements, e.g., resistive-switching memory (RSM) elements, magnetic-tunnelling memory (MTM) elements, ferroelectric memory (FM) elements, and other memory elements.…”