Feasibility Analysis of Introducing Negative Number to Improve Spiking Neural Network

Hanfei Liu

doi:10.54097/81wxy076

Authors

Hanfei Liu

DOI:

https://doi.org/10.54097/81wxy076

Keywords:

Spiking neural network, Negative number, Ternary Leaky Integrate-and-Fire.

Abstract

The aim of this paper is to investigate the feasibility of adding negative numbers to the traditional Leaky Integrate-and-Fire (LIF) for Spiking Neural Network (SNN). Currently the mainstream Spiking neuron is LIF. It is well utilized in some spiking models, but the outputs of these models can only be 0 and 1, implying that they somehow reduce the information transfer represented by negative numbers. This implies that the current SNN research on introducing negative numbers is inadequate. In this paper, by borrowing the idea of Hyperbolic tangent function and the phenomenon of biological neurons transmitting inhibitory and excitatory transmitters, the Ternary Leaky Integrate-and-Fire model (TLIF) which can output 0, -1 and 1 is proposed by improving the LIF model. And by designing a simple recognition model and comparing the recognition accuracy of LIF and TLIF models under the same dataset training, it is found that TLIF can have a better recognition effect in dealing with some datasets, but its performance is worse than LIF when facing more complex data. As the result, the introduction of negative numbers is feasible for SNNs, but further research is needed for the specific improvement methods. improvement methods and neural network design need further research.

Downloads

Download data is not yet available.

References

[1] Chen G Y, Peng P X, Li G Q, Tian Y H. Training full spike neural networks via auxiliary accumulation pathway[J]. arXiv:2301.11929, 2023.

[2] Liu K H, Zhong P S, Xu D F, Xia Q, Liu M. Corrective unit based on hyperbolic tangent function[J]. Computer Integrated Manufacturing Systems, 2020(01): 145-151.

[3] Liu H, Chai H F, Sun Q, Yun X, Li X. Research Status and Application Progress of Spiking Neural Networks[J]. Engineering Science, 2023, 06: 61-79.

[4] Taha I B M, Ibrahim S, Mansour D E A. Power Transformer Fault Diagnosis Based on DGA Using a Convolutional Neural Network with Noise in Measurements[J]. IEEE Access, 2021, 9: 111162-111170.

[5] Zhou Z K, Zhu Y S, He C, Wang Y W, Yan S C, Tian Y H, Li Y. Spikformer: When spiking neural network meets transformer[J]. arXiv:2209.15425, 2022.

[6] Faizan M, Loan A S, Alkhammash I H. Highly scalable In0.95Ga0.05As based controllable leaky-integrate-fire neuron for high performance spiking neural network and applications[J]. Physica Scripta, 2025, 100(2): 025934-025934.

[7] Garg S, Chakraborty S. Neuroscience inspired neural operator for partial differential equations[J]. Journal of Computational Physics, 2024, 515: 113266.

[8] Allan F. Nerve cell function and synaptic mechanisms[J]. Anaesthesia & Intensive Care Medicine, 2022, 23(3).

[9] Very-Large-Scale Integration Research; New Very-Large-Scale Integration Research Findings from University of Grenoble Alpes Outlined (Efficient Decompression of Binary Encoded Balanced Ternary Sequences) [J]. Electronics Newsweekly, 2019.

[10] Zhou C L, Yu L T, Zhou Z K, Zhang H, Ma Z Y, Zhou H H, Tian Y H. Spikingformer: Spike-driven Residual Learning for Transformer-based Spiking Neural Network[J]. arXiv:2301.11954, 2023.

[11] Kaushik R, Akhilesh J, Priyadarshini P. Towards spike-based machine intelligence with neuromorphic computing[J]. Nature, 2019, 575(7784): 607-617.

[12] Fang W, Chen Y, Ding J, Chen D, Yu Z, Zhou H, Tian Y, et al. Spiking jelly [EB/OL]. https://github.com/fangwei123456/spikingjelly, 2022.