Enhanced Monotonic Activation Function in Convolutional Neural Network for Multiclass EEG Signal Classification

Abstract

AbstractActivation function has a significant role in the learning process of convolutional neural network (CNN). It helps in determining the output as well as normalizing the output of the neural network. Hence, designing an efficient activation function has become the recent research trend among the deep learning fraternity. Novelty in this research is developing a new activation function, Algebraic Linear Unit (ALU), which improves the performances of the CNN. Also, they reduce computation time considerably. The performance of the proposed method is analyzed with the acquired dataset and three benchmark dataset in terms of accuracy, precision, recall, and F1 score. The extensive experiments on three standard datasets and acquired dataset infer that ALU outperforms other classical activation functions with the accuracy of 86.34%, 88.91%, 89.45%, and 83.56%, respectively. As the dataset considered for experimentation are of high volume, cloud storage would be appropriate in handling these type of data. Concurrently, the ALU is also compared with other classical monotonic activation functions like Leaky Rectified Linear Unit (L-ReLU), Softplus, Exponential Linear Unit (ELU), and Scaled Exponential Linear Unit (SELU). This proposed function can be applied to classify the needs of the paralytic people based on their visual stimuli-based EEG signals.KeywordsActivation functionEEGBCIConvolutional neural networkALUCloud storage