A Curiosity-Based Learning Method for Spiking Neural Networks.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Spiking Neural Networks (SNNs) demonstrate strong performance but face challenges in real-time application due to computational demands and complex optimization.
• Biological brains utilize curiosity as an efficient mechanism for active knowledge acquisition.

Purpose of the Study:

• To develop a curiosity-based Spiking Neural Network (CBSNN) model that enhances learning efficiency and reduces computational cost.
• To investigate the efficacy of integrating a curiosity mechanism into SNNs for improved performance.

Main Methods:

• The CBSNN model employs a four-stage learning process incorporating biologically plausible plasticity principles.
• Novelty estimations guide the network to prioritize learning informative samples, dynamically updating estimations.
• Periodic retraining of all samples prevents overfitting and forgetting, ensuring robust learning until convergence.

Main Results:

• The CBSNN model achieved 98.55% accuracy on the MNIST dataset, surpassing the Voltage-driven Plasticity-centric SNN (VPSNN) while using only 54.95% of its computation cost.
• Similar performance improvements were observed across various datasets including Iris, NETtalk, Fashion-MNIST, and CIFAR-10.
• Experimental analysis confirmed that curiosity-based learning theory significantly enhances SNN efficiency.

Conclusions:

• This study presents the first practical integration of the curiosity mechanism with SNNs.
• The proposed CBSNN model offers a viable solution for efficient and accurate SNNs, enabling realistic applications within the von Neumann framework.
• The findings pave the way for more efficient and applicable SNNs in specialized tasks.