A power-adaptive neuron model and circuit implementation

Abstract

Regarding the performance degradation and battery life issues faced by mobile smart devices during low energy supply. This article thoroughly explores the strategies employed by biological neurons to stabilize spike emission frequency and reduce power consumption during low energy supply, as well as the characteristics of myelin sheath in reducing power consumption. A power-adaptive neuron model and its corresponding power-adaptive neuron circuit system (PANCS) are proposed, which adaptively adjust power consumption according to energy supply conditions. Simulation and practical experiments both indicate that PANCS has acquired power-adaptive adjustment capability (PAAC), maintaining stable spike emission frequency when the system is under insufficient energy supply. This ability increases with the degree of myelination of PANCS. Power consumption analysis indicates that both PAAC and myelination lead to a reduction in power consumption for PANCS when energy supply is insufficient. Noise experiments demonstrate that the efficacy of PAAC entails sacrificing the robustness of PANCS, and myelination cannot reverse the decrease in robustness. Research findings of this paper endow neural morphology networks with the ability to adaptively adjust power consumption according to energy supply conditions to cope with extreme situations, providing new insights for the development of AI.