Reconfigurable Two-Dimensional Activation Neuron Device.

Abstract

Realizing energy-efficient and compact neuron devices, as well as integrating digital and analog signal processing capabilities within such a device, facilitates implementation of neural network edge computing but remains challenging. Herein, we develop a reconfigurable two-dimensional (2D) activation neuron featuring digital-analog mixed processing capabilities by 2D heterostructure. The band-to-band tunneling in the homojunction by gate-drain overlap design within the 2D channel enables the activation neuron to dynamically reconfigure subthreshold swings and threshold voltage, which is verified experimentally and by technology computer-aided design simulations and a rigorous analytical model. Consequently, the 2D activation neuron accurately replicates standard <i>Softplus</i> activation function while maintaining an ultralow leakage power of 10 pW. In neural network training, the activation neuron achieves accuracies exceeding 98%. These findings provide a potentially effective route for energy-efficient implementation of neural networks and edge computing systems.