Abstract:In analog in‐memory computing systems based on nonvolatile memories such as resistive random‐access memory (RRAM), neural network models are often trained offline and then the weights are programmed onto memory devices as conductance values. The programmed weight values inevitably deviate from the target values during the programming process. This effect can be pronounced for emerging memories such as RRAM, PcRAM, and MRAM due to the stochastic nature during programming. Unlike noise, these weight deviations do not change during inference. The performance of neural network models is investigated against this programming variation under realistic system limitations, including limited device on/off ratios, memory array size, analog‐to‐digital converter (ADC) characteristics, and signed weight representations. Approaches to mitigate such device and circuit nonidealities through architecture‐aware training are also evaluated. The effectiveness of variation injection during training to improve the inference robustness, as well as the effects of different neural network training parameters such as learning rate schedule, will be discussed. In nonvolatile memory‐based analog in‐memory computing systems, variations during device programming can cause neural‐network inference accuracy to degrade since the stored weights will differ from those in the original models. Herein, the performance of deep neural‐network models is investigated against this effect under realistic system limitations, including limited device on/off ratios, memory array size, circuit characteristics, and signed weight representations.

Origin and suppressing methodology of intrinsic variations in metal-oxide RRAM based synaptic devices

Robustness of Neuromorphic Computing with RRAM-based Crossbars and Optical Neural Networks.

Reliable Memristor-based Neuromorphic Design Using Variation- and Defect-Aware Training

Suppress Variations Of Analog Resistive Memory For Neuromorphic Computing By Localizing V-O Formation

Interface and Doping Engineering of HfO2 Based Multi-Level RRAM: Towards Synaptic Simulation for Neuromorphic Computation.

Atomic-Device Hybrid Modeling of Relaxation Effect in Analog RRAM for Neuromorphic Computing

Metal Oxide Resistive Random Access Memory Based Synaptic Devices for Brain-Inspired Computing

RRAM based synaptic devices for neuromorphic visual systems

Interface-engineered reliable HfO2-based RRAM for synaptic simulation

Modeling disorder effect of the oxygen vacancy distribution in filamentary analog RRAM for neuromorphic computing

Statistical Modeling of Time-Dependent Post-Programming Conductance Drift in Analog RRAM

Stochastic Learning in Oxide Binary Synaptic Device for Neuromorphic Computing.

Investigation of Conductance Fluctuations of Analog RRAM During Weight Update Process with Kinetic Monte Carlo Simulation

Variation-aware, Reliability-Emphasized Design and Optimization of RRAM Using SPICE Model

Electrical Properties and Biological Synaptic Simulation of Ag/MXene/SiO2/Pt RRAM Devices

A neuromorphic visual system using RRAM synaptic devices with Sub-pJ energy and tolerance to variability: Experimental characterization and large-scale modeling

Device Variation Effects on Neural Network Inference Accuracy in Analog In‐Memory Computing Systems

Bionic Sypantic Application of OxRRAM Devices

Impact of RTN and Variability on RRAM-Based Neural Network

Enhance the Robustness to Time Dependent Variability of ReRAM-Based Neuromorphic Computing Systems with Regularization and 2R Synapse

Oxide-based RRAM: Unified microscopic principle for both unipolar and bipolar switching