Stochastic Resonance Modeling of Floating Gate-Based Neurons in Summing Networks for Accurate and Energy-Efficient Operations

The stochastic resonance (SR) effect is modeled and simulated for a summing network with floating-gate (FG)-based leaky integrate-and-fire (LIF) neurons. The electron injection stochasticity, inherent in the FG-based device, is used as the stochasticity source, eliminating the need for external noise generator circuits. To evaluate the performance of the summing network, in addition to the detection accuracy, the concept of energy efficiency is introduced. The various types of noise and variation including external noise, interdevice variation, and intradevice noise are compared by simulation. The intradevice noise, which originates from the electron injection stochasticity, shows the advantage in the tradeoff between accuracy and energy efficiency. In addition, the time evolution of the threshold voltage (
${V} _{\text {TH}}$
) distributions of the network is simulated and visualized. The steady-state
${V} _{\text {TH}}$
close to
${V} _{\mathrm {TH\_{}spike}}$
and the nonlinearity of the tunneling electron injection are confirmed to enhance the detection capability. Finally, the response time of the networks is evaluated. Compared to individual neurons, the summing network with FG LIF neurons demonstrates a 50 times faster response time to detect weak and fast signals.