Low Power Ternary State Channel Computing-in-Memory Transistor for Federated Learning.

Abstract

Federated learning is a communication collaborative learning architecture that protects personal privacy. Utilization of the ternary weight changes can efficiently reduce the communication workload and energy consumption of federated learning, which is beyond the realization of most single transistors. Here, we demonstrate a ternary state channel computing-in-memory transistor that can generate three conductivity states at a minimum ternary voltage of 5 mV for low-power computing and can distinguish the direction of weight changes for federated learning. The transistor has a relatively gentle conductivity state, where the current remains almost constant as the number of voltage pulses increases. Detection of this relatively gentle conductivity state can program the change of ternary weights in a single transistor. Based on these characteristics, the total communication bits were reduced by 83.3% in a custom federated learning task. The ternary state channel transistor shows potential as a basic hardware unit for ternary computing.