Training Spiking Neural Networks for Cognitive Tasks: A Versatile Framework Compatible With Various Temporal Codes.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Supervised learning in spiking neural networks (SNNs) is effective but faces challenges like learning stability and gradient issues.
• Trainable SNNs are valuable for both engineering applications and theoretical neuroscience research.
• Existing methods often lack the biological realism needed to fully understand neural computation.

Purpose of the Study:

• To propose a modified SpikeProp learning algorithm for enhanced learning stability in SNNs.
• To introduce a framework that supports diverse network structures and coding schemes.
• To incorporate biologically realistic features into SNNs for improved neural computation modeling.

Main Methods:

• Developed a spike gradient threshold rule to mitigate the gradient exploding problem during SNN training.
• Implemented regulation rules for firing rates and connection weights to control network activity.
• Integrated biologically plausible features like lateral connections, complex synaptic dynamics, and sparse activities.

Main Results:

• Successfully trained SNNs for tasks including handwritten digit recognition, spatial coordinate transformation, and motor sequence generation using temporal codes.
• Observed emergent biologically relevant features in the trained model, such as selective activity, excitatory-inhibitory balance, and weak pairwise correlations.
• Demonstrated the framework's versatility in modeling various neural behaviors and studying underlying computational mechanisms.

Conclusions:

• The modified SpikeProp algorithm offers improved stability and flexibility for SNN training.
• The framework facilitates the creation of biologically realistic SNNs capable of complex computations.
• This approach provides a valuable tool for investigating neural function and computational principles in neuroscience.