Research on learning mechanism designing for equilibrated bipolar spiking neural networks
- PDF / 2,175,884 Bytes
- 27 Pages / 439.37 x 666.142 pts Page_size
- 1 Downloads / 193 Views
Research on learning mechanism designing for equilibrated bipolar spiking neural networks Xu Yang1 · Jiajun Lin1 · Wenhao Zheng1 · Jinfeng Zhao1 · Mengyao Ji1 · Yunlin Lei1 · Zenghao Chai1
© Springer Nature B.V. 2020
Abstract Artificial Intelligence (AI) has become very popular due to both the increasing demands from applications and the booming of computer techniques. Spiking Neural Network (SNN), as the third generation of Artificial Neural Network, receives more and more attention in the field of AI. With the high similarity to biological neural network, SNN has the potential to break through the barriers of strong AI. However, the using of SNNs on practical scenarios is rather limited, as a result of the lack of high efficient learning algorithms. Nowadays, learning methods of SNNs are designed mainly based on previous biological discoveries. The fact that there are both excitatory neurons and inhibitory neurons in the biological neural network has stimulated the motive of this research. The existence of inhibitory neurons could strengthen the self-regulation ability of neural networks and improve learning efficiency. Inspired by the ancient Chinese “Yin and Yang” Theory, we first presented our effort at constructing SNN structure with equilibrated excitatory neurons and inhibitory neurons. Then an ensemble learning optimized supervised learning method is designed and tailored for this SNN structure. Experiments are conducted using MNIST data sets, and results show that, with the designed learning mechanism, our equilibrated bipolar SNN structure could gain reasonable accuracy with much more compact structure and much more sparse synapse connections. Keywords Artificial intelligence · Spiking neural network · Supervised learning method · Ensemble learning · Excitatory and inhibitory
1 Introduction Artificial Neural Network (ANN) is a kind of information processing model inspired and abstracted from biological neural network. Large numbers of computational/storage nodes, which could be abstracted as different kinds of activation functions, are connected together, using different connection topology and different connection weight to represent different real world logic relations. Recent year has seen numerous applications of ANNs in the * Xu Yang [email protected] 1
School of Computer Science and Technology, Beijing Institute of Technology, Beijing, China
13
Vol.:(0123456789)
X. Yang et al.
domain of signal processing, pattern recognition and trend prediction, and so on, with great success. Research of ANNs has become the hot-spot in AI field. As we said, traditional ANN is inspired and abstracted from biological neural network, but only from the neural network structure. The neuron model used in traditional ANN is different from the one in biological neural network, so does the information processing method. This fact of traditional ANN has led to its limitations: 1. Poor generalization ability and accuracy when trained with small amount of data (Nguyen et al. 2015); 2. Poor energy efficie
Data Loading...
