Few-Shot Learning in Spiking Neural Networks by Multi-Timescale Optimization.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• Few-shot learning in spike-based machine learning remains a challenge due to limited task-specific prior knowledge.
• Spiking neural networks (SNNs) struggle with rapid concept learning from minimal data.
• Existing learning-to-learn (L2L) approaches for SNNs do not adequately address diverse neural dynamics on multiple timescales.

Purpose of the Study:

• To develop a novel multi-timescale optimization (MTSO) framework for SNNs.
• To enable SNNs to effectively learn from few examples by leveraging diverse temporal dynamics.
• To improve the adaptability and prior knowledge acquisition capabilities of SNNs.

Main Methods:

• Introduced an adaptive-gated Long Short-Term Memory (LSTM) to manage distinct neural dynamics timescales.
• Implemented a surrogate gradient online learning (SGOL) algorithm for fast, short-term knowledge acquisition.
• Optimized the LSTM guidance process for slow, long-term integration of knowledge and prior formation.

Main Results:

• The MTSO framework effectively accommodates both short-term learning and long-term knowledge evolution in SNNs.
• The SGOL algorithm, guided by LSTM, facilitates rapid gradient updates with adaptive learning rates and weight decay.
• Experimental results show promising performance of the proposed framework on few-shot learning tasks.

Conclusions:

• Collaborative optimization of multi-timescale neural dynamics significantly enhances SNN performance in few-shot learning scenarios.
• The MTSO framework offers a new direction for advancing SNNs in rapid learning and knowledge integration.
• Leveraging diverse temporal dynamics is crucial for improving the efficiency and effectiveness of SNNs.