Hardware-based spiking neural network architecture using simplified backpropagation algorithm and homeostasis functionality

Bio-inspired hardware-based spiking neural networks (SNNs) has been suggested as a promising computing system with low power consumption and parallel operation. We propose the supervised on-chip training method approximating the backpropagation algorithm and the pulse scheme applicable to the hardwa...

Celý popis

Uložené v:
Podrobná bibliografia
Vydané v:Neurocomputing (Amsterdam) Ročník 428; s. 153 - 165
Hlavní autori: Kim, Jangsaeng, Kwon, Dongseok, Woo, Sung Yun, Kang, Won-Mook, Lee, Soochang, Oh, Seongbin, Kim, Chul-Heung, Bae, Jong-Ho, Park, Byung-Gook, Lee, Jong-Ho
Médium: Journal Article
Jazyk:English
Vydavateľské údaje: Elsevier B.V 07.03.2021
Predmet:
Hardware-based neural networks
Homeostasis
On-chip training
Spiking Neural Networks (SNNs)
Supervised learning
Synaptic devices
Homeostasis
Hardware-based neural networks
On-chip training
Spiking Neural Networks (SNNs)
Synaptic devices
Supervised learning
ISSN:0925-2312, 1872-8286
On-line prístup:Získať plný text
Tagy: Pridať tag
Žiadne tagy, Buďte prvý, kto otaguje tento záznam!
Popis
Shrnutí:Bio-inspired hardware-based spiking neural networks (SNNs) has been suggested as a promising computing system with low power consumption and parallel operation. We propose the supervised on-chip training method approximating the backpropagation algorithm and the pulse scheme applicable to the hardware-based SNNs with the low memory dependency. The performance evaluation through the MNIST data set classification shows that the proposed system achieves a similar recognition rate compared to that of the software-based network. In addition, we also propose novel homeostasis functionality using bias synapse to achieve high performances. The homeostasis functionality well regulates the firing rate of the neurons and improves the recognition rate. The TFT-type flash memory cells are used as synaptic devices. A fully connected two-layer neural network with non-leaky integrate-and-fire (I&F) neurons is used in the simulation. We then investigate the effect of the variation of the hardware-based network on the recognition rate. The simulation results show that the proposed system is resistant to weight variation because on-chip training is adopted.
ISSN:0925-2312
1872-8286
DOI:10.1016/j.neucom.2020.11.016