9 June 2020 in IEEE Transactions on Electron Devices
In-memory computing based on memristor is a promising solution to accelerate on-chip deep neural networks. Concerning the nonideal factors of the device analog behaviors, binary neural network (BNN) with ±1 weight and 0/+1 neuron is an alternative route to better employ the binary memristors with high technical maturity. In this article, we demonstrate a select column scheme for the BNN inference accelerator. By incorporating the performance of a W/AlO x /Al 2 O 3 /Pt memristor, the BNN shows high robustness and achieves a high recognition accuracy of 98.31% on the MNIST data set. We further propose a bit error model to investigate the impact of device error on recognition accuracy. The effects of device variation and nonideal input are also considered. The results show that the accuracy keeps satisfying (more than 90%) even under a high invalid rate, high input swing (0.21 standard deviation), and 7% salt noise in binary input, respectively, whereas the device variation has a trivial effect on the network.
3 September 2020 in Advanced Intelligent Systems
Recently, due to the development of big data and computer technology, artificial intelligence (AI) has received extensive attention and made great progress. Edge intelligence pushes the computing center of AI from the cloud to individual users, making AI closer to life, but at the same time puts forward higher requirements for the realization of hardware, especially for edge acceleration. Taking convolutional neural networks (CNNs) as an example, which show excellent problem‐solving capabilities in different fields of academia and industry, it still faces issues of enormous computing volume and complex mapping architecture. Based on the computing‐in‐memory property and parallel multiply accumulate (MAC) operations of the emerging nonvolatile memristor arrays, herein the recent research progress of the edge intelligence memristive convolution accelerator is summarized. Furthermore, aiming at improving memristive convolutional accelerators, two potential optimization schemes are also discussed: The compression methods represented by quantization show great potential for static image processing, and the combination of a CNN with a long short‐term memory (LSTM) neural network makes up for the CNN's shortcomings of dynamic target processing. Finally, the future challenges and opportunities of edge intelligence accelerators based on memristor arrays are also discussed.
10 March 2026 in IEEE Journal of Solid-State Circuits
Ventricular arrhythmias (VAs), particularly ventricular tachycardia and fibrillation, remain a leading cause of sudden cardiac death. Implantable cardioverter defibrillators (ICDs) can deliver life-saving shocks, but inappropriate shocks degrade patient outcomes and device efficiency. While deep learning has improved VA detection accuracy, it lacks uncertainty quantification (UQ), limiting clinical acceptability. This article introduces a Bayesian convolutional neural network (Bayes-CNN) accelerator, designed for real-time intracardiac electrogram classification within ICDs. It combines a fully digital 1-D convolution pipeline for feature extraction with a mixed-signal, compute-in-memory (CIM) Bayesian classifier. By embedding a 360-fJ/Sample Gaussian random number generator within the CIM tile, this 0.816-mm ${}^{\mathrm {\mathbf {2}}}$ design delivers 350-pW VA detection with UQ, enabling a selective prediction strategy that reduces inappropriate shocks without compromising reliability.
21 July 2025 in IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems
Compute-in-memory accelerators built upon nonvolatile memory devices excel in energy efficiency and latency when performing deep neural network (DNN) inference, thanks to their in-situ data processing capability. However, the stochastic nature and intrinsic variations of nonvolatile memory devices often result in performance degradation during DNN inference. Introducing these nonideal device behaviors in DNN training enhances robustness, but drawbacks include limited accuracy improvement, reduced prediction confidence, and convergence issues. This arises from a mismatch between the deterministic training and nondeterministic device variations, as such training, though considering variations, relies solely on the model’s final output. In this work, inspired by control theory, we propose negative feedback training (NeFT)—a novel concept supported by theoretical analysis—to more effectively capture the multiscale noisy information throughout the network. We instantiate this concept with two specific instances, oriented variational forward (OVF) and intermediate representation snapshot (IRS). Based on device variation models extracted from measured data, extensive experiments show that our NeFT outperforms existing state-of-the-art methods with up to a 45.08% improvement in inference accuracy while reducing epistemic uncertainty, boosting output confidence, and improving convergence probability. These results underline the generality and practicality of our NeFT framework for increasing the robustness of DNNs against device variations. The source code for these two instances is available at https://github.com/YifanQin-ND/NeFT_CIM.