Fig. 4: Demonstration of multiply accumulate operation using memtransformer and BCNN for handwritten digit recognition.

a Schematics of memtransformer based on MSHE. b \({V}^{{out}}\) as a function of \({V}^{{in}}\) with positive and negative weights in single memtransformer device. c The measured values \({V}_{{measured}}^{{out}}\) as a function of their calculated values \({V}_{{ex}{pected}}^{{out}}(={w}_{1}{V}_{1}^{{in}}+{w}_{2}{V}_{2}^{{in}})\) in array comprised by two memtransformer devices. The weights are denoted as (\({{{{{{\rm{w}}}}}}}_{1}^{+}\),\(\,{{{{{{\rm{w}}}}}}}_{2}^{+}\)), (\({{{{{{\rm{w}}}}}}}_{1}^{+}\),\(\,{{{{{{\rm{w}}}}}}}_{2}^{-}\)), (\({{{{{{\rm{w}}}}}}}_{1}^{-}\),\(\,{{{{{{\rm{w}}}}}}}_{2}^{+}\)), (\({{{{{{\rm{w}}}}}}}_{1}^{-}\),\(\,{{{{{{\rm{w}}}}}}}_{2}^{-}\)). The inset is schematic of two-memtransformer-based circuits. d Structure of five-layer binary neural network used for MNIST image recognition, with two convolutional layers, one pooling layer and two fully connected layers. e The corresponding memtransformer-based arrays of the convolutional kernels in (d). f Ten thousand MNIST handwritten-digit (0,1,…,9) images are classified with a 96.7% accuracy, which is comparable to recognition accuracy of state-of-the-art binary neural network. Recognition accuracy of the ten different output digits is classified.