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.

A New Insight and Modeling of Pulse-to-Pulse Variability in Analog Resistive Memory for On-Chip Training

The Impact of Non-linear NVM Devices on In-Memory Computing

Mitigating RC-Delay Induced Accuracy Loss in Analog In-Memory Computing: A Non-Compromising Approach

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

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

Circuit Modeling for RRAM-Based Neuromorphic Chip Crossbar Array with and Without Write-Verify Scheme.

Time-dependent variability in RRAM-based analog neuromorphic system for pattern recognition

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

A Compact Model of Analog RRAM With Device and Array Nonideal Effects for Neuromorphic Systems

The Impact of Endurance Degradation in Analog RRAM for In-Situ Training

Intra-array Non-Idealities Modeling and Algorithm Optimization for RRAM-based Computing-in-Memory Applications

Early-Stage Fluctuation in Low-Power Analog Resistive Memory: Impacts on Neural Network and Mitigation Approach

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

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

Characterizing Endurance Degradation of Incremental Switching in Analog RRAM for Neuromorphic Systems

A Compact Model of Analog RRAM Considering Temperature Coefficient for Neural Network Evaluation

A Compact Model of Analog RRAM for Neuromorphic Computing System Design

Relaxation Signal Analysis and Optimization of Analog Resistive Random Access Memory for Neuromorphic Computing

Compensation Architecture to Alleviate Noise Effects in RRAM-based Computing-in-memory Chips with Residual Resource

Reliability-Aware Training and Performance Modeling for Processing-In-Memory Systems

Reliability Perspective on Neuromorphic Computing Based on Analog RRAM