Trajectory Prediction System Utilizing the Intrinsic Long-Term and Short-Term Plasticity in CT-FeFET

Trajectory prediction is a vital function in the auto-driving field which relies on both historical information and current inputs to make forecasts. Continuous attractor neural network (CANN) with dynamic synapses is one of the typical algorithms for conducting prediction tasks. However, current devices are limited to separately handling static weight storage and dynamic modulation, leading to considerable resource consumption in the hardware implementation of CANN. Here, we integrate long-term and short-term plasticity using charge-trapping ferroelectric FETs (CT-FeFETs) to realize dynamic synaptic units. The long-term weights are stored in the ferroelectric domain, while the short-term weights are dynamically operated in a CT domain, eliminating the external caching process. Based on the intrinsic long-term and short-term plasticity of CT-FeFETs, historical information interacts with the inputs in situ, facilitating a dynamic in-memory computing (IMC) that enables real-time prediction. Moreover, we introduced CT-FeFET into both feedforward networks and CANN to validate prediction performance, and confirmed that a larger dynamic range and smaller time constant in the short-term dynamics enable the system to match a broader speed range for predictive tracking. This work provides a novel way of performing interacting tasks with dynamic IMC technology.