An STDP training algorithm for a spiking neural network with dynamic threshold neurons.

Area of Science:

• Computational Neuroscience
• Machine Learning
• Artificial Intelligence

Background:

• Spiking Neural Networks (SNNs) offer bio-realistic computation but face training challenges.
• Existing Spike Timing Dependent Plasticity (STDP) rules are often limited in scope for complex network architectures.
• Supervised learning in SNNs requires adaptable plasticity rules to handle network-level interactions.

Purpose of the Study:

• To propose a novel supervised training algorithm for Spiking Neural Networks.
• To extend the Spike Timing Dependent Plasticity (STDP) learning rule for multi-synaptic connections and axonal delays.
• To evaluate the algorithm's performance on benchmark datasets and investigate neuron dynamics.

Main Methods:

• Development of a modified STDP learning rule supporting supervised, local, and network-level training.
• Application of the algorithm to two- and three-layer SNN architectures.
• Benchmarking using the Iris and Wisconsin Breast Cancer (WBC) datasets.
• Investigation of dynamic threshold neurons in hidden layers.

Main Results:

• The proposed supervised training algorithm effectively trains multi-layer SNNs.
• The modified STDP rule successfully accommodates multiple synaptic connections and axonal delays.
• Performance evaluation on Iris and WBC datasets demonstrates the algorithm's efficacy.
• Dynamic threshold neurons show potential for improved SNN performance.

Conclusions:

• The developed supervised training algorithm offers a robust method for training SNNs.
• The enhanced STDP rule provides a flexible framework for complex SNN architectures.
• This work contributes to advancing the practical application of SNNs in machine learning.