Supervised Learning Based on Temporal Coding in Spiking Neural Networks.

Area of Science:

• Computational neuroscience
• Artificial intelligence

Background:

• Gradient descent is effective for artificial neural networks (ANNs).
• Training spiking neural networks (SNNs) is challenging due to non-linearities and discrete spike events.
• Existing methods often use rate-based coding, which differs from biological neuron behavior.

Purpose of the Study:

• To adapt gradient descent training for feedforward spiking neural networks.
• To explore temporal coding schemes for information representation in SNNs.
• To demonstrate efficient training of SNNs on complex tasks.

Main Methods:

• Utilized a temporal coding scheme where information is encoded in spike timing.
• Showed that the network's input-output relation is differentiable almost everywhere.
• Demonstrated piecewise linearity after a variable transformation, enabling gradient-based optimization.
• Trained the spiking network on the permutation invariant MNIST task.

Main Results:

• Successfully adapted gradient descent for training spiking neural networks with temporal coding.
• Achieved efficient training on the permutation invariant MNIST task.
• Developed SNNs that exhibit sparse spiking, unlike conventional ANNs or rate-based SNNs.

Conclusions:

• Gradient descent training methods are directly applicable to feedforward SNNs using temporal coding.
• This approach offers a new way to control SNN behavior with realistic temporal dynamics.
• Enables SNNs to process complex temporal information patterns effectively.