What does this research mean for the field?

In analog spiking neural networks trained via unsupervised STDP, simplifying or omitting random noise modeling improves learning performance, and neglecting noise during training causes only a minor accuracy drop (0.2% to 4.5%) when reintroduced during testing. Novelty: ClaimNovelty.INCREMENTAL. Consensus alignment: ConsensusAlignment.NEUTRAL.

December 23, 2024Open Access

Analog Spiking Neuron Model for Unsupervised STDP-based learning in Neuromorphic Circuits

YBYannaël BossardUniversité Paris-Saclay ZJZalfa JouniCentre National de la Recherche Scientifique SWSiqi WangChongqing University

Key Points

Key points are not available for this paper at this time.

Abstract

Neuromorphic computing has emerged as a promising solution to meet the computational demands of artificial intelligence and enhance the energy efficiency through spiking neural networks (SNNs). This paper models a previously published biomimetic electronic neuron (eNeuron), focusing on the effects of its random noise within an analog SNN framework. The analog SNN is trained via Spike-Timing-Dependent Plasticity (STDP) using unsupervised learning method. Its computational efficiency is evaluated against two benchmarks. Preliminary results have demonstrated that the analog SNN with either simplified or omitted random noise modeling perform better than those with full noise modeling. This highlights the significant impact that random noise of transistors in the eNeuron has on STDP learning processes. However, if random noise is neglected during the training phase, the reintroduction of noise modeling in the testing phase causes an average accuracy drop of only 0.2% for the simplified model of random noise and 4.5% for the full model.

AI에게 질문

Bookmark

View Full Paper