کنفرانس مهندسی برق ایران

صفحه اصلی / سی و سومین کنفرانس بین المللی مهندسی برق

Classifier Fusion Based on Extracted Features Using a Spiking Neural Network from Handwritten Digits

نویسندگان :

Ali Gholamzade Fard Kazzazi¹ Malihe Nazari² Fariba Bahrami³

1- دانشگاه صنعتی امیرکبیر 2- دانشگاه تهران 3- دانشگاه تهران

کلمات کلیدی :

spiking neural network،spike timing dependent plasticity،unsupervised learning،feature extraction،classification،classical classifiers،classifier fusion

چکیده :

The third generation of neural networks, known as spiking neural networks (SNNs), are capable of solving all problems that traditional networks can solve, and computationally, they can perform even more powerfully. Spiking neurons are closer to biological reality. For these reasons, these networks have gained significant attention in recent years. When using these networks for tasks such as pattern recognition or classification, there is no precise method for data classification. In previous works that employed spiking neural networks for classification, each approach generally utilized the unsupervised learning mechanisms that existed in these networks to classify the data through different techniques. Due to the weaknesses in the classification layer of spiking neural networks, we turned to the use of the firing rates of spiking neurons to extract features, which were then passed to classical classifiers. When we used spiking neurons for classification, we achieved an accuracy of 80.17%, and when we added a classical classifier in the third layer of the network, the accuracy increased to 84.48%. Based on the results obtained, the use of a classical classifier layer improved the network's accuracy. Additionally, it increased the execution speed compared to the case where a spiking neuron layer was used in the classification layer, and it reduced the need for extensive hyperparameter tuning of the SNN. Finally, we applied the Decision Template method for classifier fusion, which led to an accuracy of 84.87%. The results show that using classifier fusion methods improves the performance of the network.