Heterogeneous Ensemble-Based Spike-Driven Few-Shot Online Learning.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking neural networks (SNNs) offer energy efficiency but lag in few-shot learning compared to ANNs.
• Existing SNN few-shot models lack robustness and theoretical grounding in spatiotemporal dynamics.
• Bridging the gap in few-shot learning performance is crucial for SNNs' practical application.

Purpose of the Study:

• To propose a novel spike-based framework, HESFOL, for robust few-shot online learning.
• To leverage entropy theory within a recurrent SNN architecture for gradient-based learning.
• To enhance the spatiotemporal dynamics and machine learning theory basis of SNN few-shot learning.

Main Methods:

• Developed a heterogeneous ensemble-based spike-driven framework (HESFOL).
• Integrated entropy theory to establish a gradient-based few-shot learning scheme.
• Utilized recurrent SNN architecture for processing spatiotemporal data.

Main Results:

• HESFOL demonstrated improved accuracy and robustness in few-shot classification tasks (spiking patterns, Omniglot).
• The framework showed effectiveness in a few-shot motor control task (end-effector).
• Validated the successful application of entropy-based machine learning in SNNs.

Conclusions:

• HESFOL effectively enhances spike-driven few-shot learning performance.
• The study highlights the merit of integrating entropy theory into SNNs for improved learning.
• Provides new perspectives for advancing neuromorphic systems through advanced machine learning theories.