Pattern Recognition of Spiking Neural Networks Based on Visual Mechanism and Supervised Synaptic Learning

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Machine learning

Background:

• Mammalian primary visual cortex neurons exhibit orientation selectivity to visual stimuli.
• Traditional deep learning models require large datasets and high computational power for image recognition.

Purpose of the Study:

• To develop a hierarchical spiking neural network (SNN) for image classification inspired by the visual cortex.
• To demonstrate the efficacy of the SNN in reducing training data requirements while maintaining classification accuracy.

Main Methods:

• A feed-forward network with orientation-selective neurons processes grayscale images.
• A supervised tempotron learning rule facilitates communication between layers.
• The network is trained and evaluated on the MNIST handwritten digit database.

Main Results:

• The SNN achieved 96% classification accuracy on the MNIST dataset.
• The model required only 2000 training samples, a significant reduction from the traditional 60000.
• The approach demonstrates biological plausibility and high classification accuracy.

Conclusions:

• The proposed brain-inspired SNN model offers an accurate and data-efficient alternative for image classification.
• This hierarchical processing approach may enable wider applications of SNNs in intelligent computing.
• The model addresses limitations of deep learning, such as large data needs and power consumption.