활성화 함수를 시간적으로 근사하기 위한 스파이킹 뉴런 모델의 노이즈 주입에 따른 학습 성능 분석

Abstract

IN spiking neural networks (SNNs), the leaky integrate-and-fire (LIF) neurons, which are widely used model, have limitations in approximating various activation functions, such as Swish and GeLU. To address this, few spikes (FS) neurons were proposed but encountered generalization challengers. This study investigate noise injection in training of FS neurons to enhance the generalization performance. Experimental results with Gaussian noise showed that there was insufficient improvement in training results. This suggests the need for training algotithms to improve the generalization performance of FS neurons.