Neuronal-Plasticity and Reward-Propagation Improved Recurrent Spiking Neural Networks | JoVE Visualize

Neuronal-Plasticity and Reward-Propagation Improved Recurrent Spiking Neural Networks

Shuncheng Jia^1,2, Tielin Zhang^1,2, Xiang Cheng^1,2

¹Research Center for Brain-Inspired Intelligence, Institute of Automation, Chinese Academy of Sciences (CASIA), Beijing, China.

Frontiers in Neuroscience|March 29, 2021

View abstract on PubMed

Summary

We introduce a novel Recurrent Spiking Neural Network (NRR-SNN) that enhances biological plausibility and computational robustness. This new model achieves higher accuracy on sequential datasets through improved neuronal plasticity and reward propagation.

Area of Science:

Computational Neuroscience
Artificial Intelligence

Background:

Spiking Neural Networks (SNNs) leverage biologically-plausible principles for efficient and robust computation, outperforming Deep Neural Networks (DNNs).
Existing SNNs benefit from natural neural network dynamics and plasticity, but further improvements are sought for complex tasks.

Purpose of the Study:

To propose a novel Recurrent Spiking Neural Network (NRR-SNN) integrating enhanced neuronal plasticity and reward propagation mechanisms.
To improve the accuracy and robustness of SNNs for sequential data processing.

Main Methods:

Introduced a special Neuronal-plasticity and Reward-propagation improved Recurrent SNN (NRR-SNN).
Incorporated an adaptive threshold with two channels as a key neuronal plasticity mechanism to boost neuronal dynamics.
Utilized global labels as rewards for gradient propagation, replacing traditional error signals.
Designed a recurrent loop with controlled sparseness for enhanced computational robustness.

Main Results:

The NRR-SNN demonstrated higher accuracy on sequential datasets, specifically the TIDigits and TIMIT datasets.
Achieved stronger robust computation capabilities compared to existing SNN models.
The biologically-plausible improvements contributed to the enhanced performance.

Conclusions:

The proposed NRR-SNN effectively integrates advanced neuronal plasticity and reward propagation for superior performance.
The model showcases the potential of biologically-inspired enhancements in SNNs for complex sequential tasks.
NRR-SNN represents a significant advancement in developing efficient and robust artificial neural networks.