Supervised Learning in Multilayer Spiking Neural Networks With Spike Temporal Error Backpropagation.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) offer low power consumption and high computational power.
• Current SNN learning algorithms primarily focus on synaptic weight adjustment, neglecting the role of synaptic delays.
• Neuroscience indicates synaptic delays are crucial for learning processes.

Purpose of the Study:

• To develop a novel gradient descent-based learning algorithm for synaptic delays in SNNs.
• To enhance the sequential learning capabilities of single spiking neurons and extend it to multilayer SNNs.
• To improve the efficiency and accuracy of SNNs in complex tasks like audiovisual pattern recognition.

Main Methods:

• A gradient descent-based algorithm was developed to modulate synaptic delays.
• The algorithm was extended to multilayer SNNs using spike temporal-based error backpropagation.
• Information encoding relies on the relative timing of neuronal spikes, with learning based on precise spike time derivatives.

Main Results:

• Significant improvements in learning efficiency and accuracy were observed on synthetic and realistic datasets.
• The proposed method outperformed existing spike temporal-based learning algorithms.
• The algorithm demonstrated superior performance in an SNN-based multimodal model for audiovisual pattern recognition.

Conclusions:

• Modulating synaptic delays is an effective strategy for enhancing SNN learning, particularly for sequential tasks.
• The proposed gradient descent-based algorithm offers a powerful new tool for SNN development.
• This approach advances the application of SNNs in complex real-world problems such as pattern recognition.