大数据时代下，提高信号处理效率至关重要。由于传统计算架构的计算设备中存储单元和计算单元相互分离，未来将面临着计算效率的限制。基于多电导态器件的阵列电路可实现全硬件卷积神经网络（CNNs），具备提高计算效率的潜力。然而，在处理大规模卷积计算时，仍存在大量器件冗余，导致计算功耗低、计算成本高。
近日，福州大学陈惠鹏教授和杨倩博士等人在Science China Materials发表研究论文，创新性地提出了一种基于忆阻器的器件级原位卷积策略：以忆阻器的导电丝、掺杂面积和极化面积等的动态变化作为卷积运算过程，单个器件的电导切换所需的时间作为计算结果，通过忆阻器独特的尖峰数字信号体现卷积计算。
本文要点
1) 通过忆阻器将复杂的模拟信号合理地编码为简单的数字信号，成功在单个器件上完成了卷积计算，这对于复杂信号处理和计算效率提高至关重要。2) 在器件级原位卷积计算的基础上，进一步实现了盲文信号的特征识别和噪声过滤。这项工作所提的基于单个忆阻器的器件级原位卷积计算，将推动具有大规模卷积计算能力的复杂CNNs的构建，促进神经形态计算领域的创新和发展。Figure 1. Basic strategy of CNNs based on in-situ convolutional memristors. (a) Traditional CNNs neuromorphic computing hardware system based on multi-channel devices. (b) A new CNNs basic strategy using the dynamic change of memristor conductive wire as the convolution operation process.Figure 2. Basic electrical performance test of convolutional memristor. (a) Typical threshold switching performance of the neuron device. (b) Distribution diagram of 50 randomly extracted threshold voltages. (c) Endurance test under 250 DC cycles. (d) Measure integrate-and-firing behavior of the neuron device. (e) The firing time of the neuron device is affected by the amplitude of stimulation pulse (2.8, 3.2, and 3.6 V). (f) The firing time is affected by the duty ratio of stimulation pulse (25%, 50% and 75%).![]()
Figure 3. The convolutional calculation of In-SCM.(a) Mathematical process of convolution calculation. (b) Encoding rules for input data matrix and weight matrix. (c) Mathematical process of different convolution calculations. (d) Resistance switching time of In-SCM corresponding to different convolution calculation results, where the figures at the same row show the mathematical convolution calculations performed by In-SCM (for example, Fig. 3c (ii) and 3d (ii) show the hardware calculation of [1, 2; 3, 4] convolved [1, 0; 0, 0]). (e) Working mechanism of In-SCM.Figure 4. Implementation of braille recognition based on In-SCM. (a) Topology structure of the perceptual computing system based on In-SCM. (b) Simulation of the closed-loop flowchart of the human sensor system. (c) Schematic of braille numbers from 0 to 9. (d) In-SCM device encodes different characteristic signals. (e) Recognition of normal braille number 0. (f) Recognition of braille number 0 under interference.Xianghong Zhang, Congyao Qin, Wenhong Peng, Ningpu Qin, Enping Cheng, Jianxin Wu, Yuyang Fan, Qian Yang, Huipeng Chen. Memristor-based in-situ convolutional strategy for accurate braille recognition. Sci. China Mater. (2024).https://doi.org/10.1007/s40843-024-3122-7
![]()
点击左下角“阅读原文”，阅读以上文章PDF原文
![]()
