Electroencephalogram (EEG) is a crucial tool in neurological diagnostics. EEG provides critical insights into brain activity, such as in applications like visual stimuli classification. EEG-based visual stimulus classification has the potential to advance brain-computer interface (BCI) technologies, enabling more effective decoding of brain activity. However, accurately classifying visual stimuli from single-trial EEG data remains challenging due to the low signal-to-noise ratio and the need for complex signal preprocessing techniques.

This thesis proposes a baseline model using the ROCKET classifier and a simplified Convolutional Neural Network (CNN) approach for visual stimulus classification. In both cases, the models are fed raw EEG data. This method aims to bypass traditional preprocessing and feature extraction methods.

The study systematically explores the influence of various input shapes on model performance and identifies the optimal configuration, Cx1xN, which achieved a competitive classification accuracy of 50.1% for a six-class task.

This method performs well as advanced combined models but with much lower computational demands. Additionally, results indicate that this approach is scalable across subjects and maintains robustness and applicability in real-time BCI applications. The findings offer valuable insights into simplifying EEG classification workflows for future BCI systems, with potential use cases in clinical and neurofeedback systems.