Direct learning-based deep spiking neural networks: a review.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience

Background:

• Spiking Neural Networks (SNNs) are brain-inspired models utilizing binary spikes.
• Their event-driven nature and temporal dynamics offer computational advantages.
• However, optimizing deep SNNs is challenging due to discontinuous spike mechanisms.

Purpose of the Study:

• To provide a comprehensive survey of direct learning-based deep SNNs.
• To categorize existing methods for training SNNs.
• To identify future research challenges and trends in SNN optimization.

Main Methods:

• Review and categorize direct learning approaches for deep SNNs.
• Focus on methods utilizing the surrogate gradient technique.
• Classify techniques based on accuracy, efficiency, and temporal dynamics utilization.

Main Results:

• Direct learning methods, particularly surrogate gradient, significantly ease SNN optimization.
• Existing works are categorized into accuracy, efficiency, and temporal dynamics improvements.
• Progress has been made in training deeper and more effective SNNs.

Conclusions:

• Direct learning offers a viable path for training deep SNNs.
• Further research is needed to address remaining challenges and explore future trends.
• SNNs hold significant potential for efficient and brain-like computation.