Scholarlink

This paper presents the Hebbian-Augmented Associative Memory (HAAM) framework, which utilizes synaptic plasticity to improve few-shot learning capabilities in Spiking Neural Networks (SNNs). By addressing the limitations of traditional learning paradigms, the HAAM framework incorporates dynamic synaptic adjustments, facilitating efficient learning from limited labeled data. Experimental evaluations conducted on benchmark datasets, such as Omniglot and MiniImageNet, demonstrate that the HAAM plasticity rule significantly outperforms leading non-pretrained SNN meta-learning methods while remaining competitive with advanced Artificial Neural Network (ANN) techniques. These findings highlight the effectiveness of biologically inspired learning mechanisms in capturing temporal dynamics and enabling rapid adaptability to novel tasks. Moreover, the HAAM framework exhibits robust generalization capabilities, positioning it as a promising solution for real-world applications characterized by data scarcity. This research contributes to the expanding domain of neuro-inspired computation, offering insights for future investigations aimed at optimizing neural architectures to enhance learning efficiency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Enhancing Few-Shot Learning in Spiking Neural Networks Through Hebbian-Augmented Associative Memory

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Weiyi Li,
  • Dongcheng Zhao,
  • Yiting Dong,
  • Guobin Shen,
  • Yi Zeng

摘要

This paper presents the Hebbian-Augmented Associative Memory (HAAM) framework, which utilizes synaptic plasticity to improve few-shot learning capabilities in Spiking Neural Networks (SNNs). By addressing the limitations of traditional learning paradigms, the HAAM framework incorporates dynamic synaptic adjustments, facilitating efficient learning from limited labeled data. Experimental evaluations conducted on benchmark datasets, such as Omniglot and MiniImageNet, demonstrate that the HAAM plasticity rule significantly outperforms leading non-pretrained SNN meta-learning methods while remaining competitive with advanced Artificial Neural Network (ANN) techniques. These findings highlight the effectiveness of biologically inspired learning mechanisms in capturing temporal dynamics and enabling rapid adaptability to novel tasks. Moreover, the HAAM framework exhibits robust generalization capabilities, positioning it as a promising solution for real-world applications characterized by data scarcity. This research contributes to the expanding domain of neuro-inspired computation, offering insights for future investigations aimed at optimizing neural architectures to enhance learning efficiency.