Analog Implementation of Sigmoidal Neuron by 2-D Semiconductive Resistor

Neuromorphic accelerators based on non-von Neumann architecture are indispensable in the future, but currently, the hardware implementation of nonlinear activation function is still too complicated. This work provided a useful activation function generator with a two-terminal MoS2/CuInP2S6 (MoS2/CIPS) resistor, which is able to output an approximate sigmoid I–V curve. This functionality was verified by the comparison between the neural network (NN) models based on the I–V function and other conventional activation functions, such as rectified linear unit (ReLU) and hyperbolic tangent (tanh), with different training datasets of Modified National Institute of Standards and Technology (MNIST), Kuzushiji-MNIST (KMNIST), and Fashion-MNIST (FMNIST). Then, the trained convolutional neural network (CNN) model was used for feature extraction of human face, demonstrating its recognition capability for various image types. Moreover, a third electrode was added as a bottom gate to explore the tunability of the MoS2/CIPS resistor based on the ferroelectric property of CIPS, which is effective in a limited gate voltage range of −1 to −0.5 V. Significantly, it empowers the device to demonstrate the synaptic plasticity of long-term potentiation (LTP). This work provided a useful sigmoid function generator for neuromorphic computing with very simple structure.