Effective Transfer Learning Algorithm in Spiking Neural Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Neuromorphic Computing

Background:

• Spiking neural networks (SNNs), the third generation of neural networks, offer high energy efficiency on neuromorphic hardware.
• Training deep SNNs necessitates extensive labeled data, which is costly and difficult to acquire, similar to traditional artificial neural networks (ANNs).
• Transfer learning, while prevalent in ANNs, has seen limited application in SNNs due to training challenges.

Purpose of the Study:

• To propose an effective transfer learning framework for deep SNNs.
• To address the challenge of limited labeled data in SNN training.
• To investigate domain-invariant representation for improved transferability.

Main Methods:

• Developed a transfer learning framework for deep SNNs utilizing domain-invariant representation.
• Analyzed the suitability of Centered Kernel Alignment (CKA) as a domain distance metric for SNNs, comparing it with Maximum Mean Discrepancy (MMD).
• Evaluated feature transferability across different network layers using benchmark datasets: Office-31, Office-Caltech-10, and PACS.

Main Results:

• Experimental results confirmed the feature transferability of SNNs.
• The proposed transfer learning framework demonstrated effectiveness in SNN applications.
• Centered Kernel Alignment (CKA) proved to be a rational and effective domain distance measurement for SNNs.

Conclusions:

• The study successfully established an effective transfer learning framework for deep SNNs.
• The findings highlight the potential of CKA in enhancing SNN transfer learning.
• This research contributes to overcoming data limitations in SNN training and promotes their wider adoption.