Design and Implementation of a Spiking Neural Network with Integrate-and-Fire Neuron Model for Pattern Recognition

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Traditional artificial neural networks (ANNs) differ from spiking neural networks (SNNs) in their temporal coding mechanisms.
• Optimizing SNN parameters based on physiological principles enhances network robustness and information processing.

Purpose of the Study:

• To implement a robust spiking neural network (SNN) inspired by the center-surround structure of retinal receptive fields.
• To evaluate the performance of the proposed SNN using the Integrate-and-Fire (IF) neuron model and time-to-first-spike coding.

Main Methods:

• The proposed SNN utilizes the Integrate-and-Fire (IF) neuron model.
• Time-to-first-spike coding is employed for network training with a novel learning method.
• The SNN is evaluated on the Iris, MNIST, and ABIDE1 datasets.

Main Results:

• Achieved 96.33% accuracy on the Iris dataset with 60 input neurons in 45 iterations.
• Reached 90.5% accuracy on MNIST with pixel input (600 neurons), improving to 95% with 210 neurons using 14 structural features.
• Attained 84.42% accuracy on the ABIDE1 dataset for autism classification using Shannon entropy, with 120 iterations.

Conclusions:

• The implemented SNN demonstrates high accuracy and efficiency across diverse datasets.
• Feature engineering significantly improves SNN performance and reduces the number of required input neurons.
• The proposed SNN shows promise for applications in pattern recognition and biomedical data analysis.