An unsupervised STDP-based spiking neural network inspired by biologically plausible learning rules and connections

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning relies heavily on labeled data, unlike human learning.
• Spiking neural networks (SNNs) with Spike-Timing-Dependent Plasticity (STDP) show promise but have performance limitations.
• Existing SNNs struggle with unsupervised learning efficiency and complex tasks.

Purpose of the Study:

• To develop an improved unsupervised learning model for SNNs inspired by biological learning mechanisms.
• To enhance the representation and learning capabilities of SNNs through adaptive plasticity.
• To accelerate and stabilize the training of unsupervised SNNs for complex pattern recognition.

Main Methods:

• Introduced an adaptive synaptic filter and adaptive spiking threshold for neuron plasticity.
• Incorporated adaptive lateral inhibitory connections for dynamic spike balance.
• Developed a samples temporal batch STDP (STB-STDP) for efficient weight updates.
• Integrated these adaptive mechanisms to create a novel unsupervised SNN training framework.

Main Results:

• Achieved state-of-the-art performance for unsupervised STDP-based SNNs on MNIST and FashionMNIST datasets.
• Demonstrated superior performance on the complex CIFAR10 dataset, marking the first application of unsupervised STDP-SNNs to this dataset.
• Showcased significant advantages in small-sample learning scenarios compared to supervised artificial neural networks (ANNs).

Conclusions:

• The proposed adaptive mechanisms and STB-STDP substantially improve unsupervised SNN training speed and performance.
• The model offers a powerful new approach for unsupervised learning in SNNs, particularly for complex and limited-data scenarios.
• This work bridges the gap between biological learning and artificial neural networks, paving the way for more efficient and human-like AI.