Power Spectral Density Based Weight Initialization Technique For Feed-Forward Neural Network

In the field of machine learning (ML), neural networks have demonstrated remarkable capabilities in delivering excellent results and solutions. Among the various features and characteristics that make neural networks unique, weight initialization is a critical step that significantly impacts outcomes. Over time, multiple methods for weight initialization have been introduced, ranging from simple random initialization to sophisticated techniques such as Nguyen-Widrow and Xavier initialization. In this article, a data-driven weight initialization technique is introduced, based on the power spectral density of electroencephalography (EEG) signals. A single-layer feedforward neural network was employed to classify three publicly available EEG datasets. The results demonstrated a significant enhancement in accuracy and an improved convergence rate, thereby optimizing the model’s performance.