Abstract
As the fundamental computational element in neuromorphic computing, the number of neurons determines the calculation power of a neuromorphic system. Therefore, emerging device like leaky-Ferroelectric FET (leaky-FeFET), which can mimic the neuron behaviors and significantly reduce the hardware cost in neuron design, has drawn increasingly attentions. However, leaky-FeFET requires additional manufacture process, which dramatically increases the cost. This paper proposes a conventional FeFET-based neuron circuit, which consists of three transistors and two resistors (3T2R), enabling both excitatory and inhibitory input connections. The leaky effect is simply realized through a periodical depolarization controlled by an external pulse signal and can be conveniently tuned by changing the frequency of the external pulse. Based on the proposed neuron, a SNN designed for the Sudoku task has been implemented with lower hardware cost. Moreover, to evaluate the benefit of the tunable leaky effect, a Figure-of-Merit (FoM A) quantifying the competitive advantage of the winner neuron is introduced. Compared with fixed leaky effect, the SNN with proposed neuron circuit shows an average of 30% improvement of the FoM, which means the proposed circuits will have higher precision under a variety of scenarios.