Journal of Nanjing University(Natural Sciences)

Figure/Table detail

Fully Analog Neural Network based on 1T1R memristor crossbar array and CMOS activation functions

Hang Zhao, Dongxingjian Yang, Cong Wang, Shijun Liang, Feng Miao

Journal of Nanjing University(Natural Sciences), 2025, 61(5): 867-878. DOI: 10.13232/j.cnki.jnju.2025.05.015

Fig.4 The flow chart to precisely tune the conductive states of the memristive device in 1T1R crossbar array

Other figure/table from this article

Fig.1 (a) 1R memristor crossbar array, (b) 1T1R memristor crossbar array
Fig.2 Optical photograph of 1T1R memristor crossbar array
Fig.3 (a) Switching curves of the 1T1R cell under different gate voltages, (b) adjusted conductance of the 1T1R cell under different gate voltages
Fig.5 Programming error of memristors
Fig.6 The conductance matrix after writing test
Fig.7 Basic structure of CMOS activation function
Fig.8 Circuit and input⁃output characteristics of CMOS pseudo ReLU activation function
Fig.9 Circuit and input⁃output characteristics of CMOS pseudo Leaky ReLU activation function
Fig.10 Circuit and input⁃output characteristics of CMOS pseudo Tanh activation function
Fig.11 Circuit and input⁃output characteristics of CMOS pseudo Sigmoid activation function
Fig.12 Circuit and input⁃output characteristics of CMOS pseudo Softmax activation function
Table 1 Comparison of power consumption and area between analog CMOS activation function circuits and high⁃speed ADC in 65 nm process node
Fig.13 The computational graph of a fully analog neural network
Fig.14 Pseudo activation function and its derivative
Fig.15 Results of training using pseudo Leaky ReLU circuits versus the standard ones
Fig.16 Results of training using pseudo Tanh circuits versus the standard ones
Fig.17 Results of training using pseudo Sigmoid circuits versus the standard ones
Fig.18 Recognition accuracy considering different weight programming errors

模态框（Modal）标题

Figure/Table detail