Power-efficient and Multiplier-less FPGA Implementation of Self-adaptive Leaky Integrate-and-Fire Neuron

Abstract

Modeling and implementation of neuron are vital for the engineering of biological neural networks and neuromorphic systems. Hodgkin-Huxley model, Izhikevich model, and leaky integrate-and-fire (LIF) model are the most commonly used models in biological neural networks. Considering the efficiency and cost of digital implementation, LIF model is a good choice to realize neural networks on chip. This paper presents a method to implement a self-adaptive LIF model on field-programmable gate array (FPGA). Based on the traditional LIF model, the calcium current of neurons is taken into consideration in our model, which enables the neuron to be self-adaptive and biologically plausible. Our efficient digital realization of LIF neuron has the highlight of (i) efficient hardware usage (654 logic elements per neuron) (ii) low power consumption (142.28 mW per neuron), and (iii) ability to realize different kinds of biological spiking behaviors. The behavior of coupled neurons model is also tested on multiple FPGA chips. The result shows that our self-adaptive LIF neuron model has a good performance of interaction between FPGA.