Effect of Different Activation Functions on EEG Signal Classification based on Neural Networks

In recent years, EEG (Electroencephalography) has grabbed significant research attention towards giving it as an input to the computing devices. To identify hidden patterns, EEG signals is not straightforward, and in order to be able to use it as a reliable input method, adequate accuracy is required. While machine learning techniques are there for numeric, categorical and similar types of data, Neural Networks are far more capable of identifying patterns and do the classification more effectively. Activation functions are integral part of the neural networks and selection of the activation function is a choice for the neural network designer. This paper showcases various activation functions and their impact on overall accuracy of classification using a dataset of 84 samples of EEG signals using ANN (Artificial Neural Network) and CNN (Convolutional Neural Network). Total 6 different activation functions considered here at all levels of ANN and CNN and respective results are analyzed. Activation functions at different layers of both ANN and CNN are compared such as fully connected layers and convolution layers. Accuracy is the measure considered for the comparison of the effectiveness of neural networks.