Robust Spike-Based Continual Meta-Learning Improved by Restricted Minimum Error Entropy Criterion.

Area of Science:

• Neuromorphic Engineering
• Machine Learning Theory
• Information Theory

Background:

• Spiking neural networks (SNNs) offer a potential solution to the high energy consumption of deep neural networks.
• Current SNNs lag behind artificial neural networks in online meta-learning performance.
• Existing spike-based meta-learning models lack focus on robust learning from spatio-temporal dynamics and advanced theory.

Purpose of the Study:

• To propose a novel spike-based framework, MeMEE (Minimum Error Entropy), for gradient-based online meta-learning in recurrent SNNs.
• To leverage entropy theory to enhance the learning capabilities of SNNs.
• To improve the accuracy and robustness of spike-based meta-learning.

Main Methods:

• Developed a recurrent SNN architecture incorporating a novel framework called MeMEE.
• Utilized entropy theory to establish a gradient-based online meta-learning scheme.
• Evaluated performance on tasks including autonomous navigation and working memory tests.

Main Results:

• The MeMEE model significantly improved the accuracy of spike-based meta-learning.
• The proposed framework demonstrated enhanced robustness in SNN performance.
• Experimental results validated the effectiveness of MeMEE on diverse tasks.

Conclusions:

• MeMEE effectively enhances accuracy and robustness in spike-based meta-learning.
• The study highlights the application of modern information theoretic learning in SNNs.
• This work offers new perspectives for integrating advanced information theory into machine learning to improve SNN performance for neuromorphic systems.