This paper explores the use of convolutional neural networks (CNNs) for feature extraction from electroencephalogram (EEG) signals. The authors aim to provide insights into creating and fine-tuning CNNs for various applications, considering the characteristics of EEG signals and various signal processing techniques. They analyze well-known CNN architectures, categorizing them into four groups: standard implementation, recurrent convolutional, decoder architecture, and combined architecture. The paper offers a comprehensive evaluation of these architectures, covering accuracy metrics, hyperparameters, and an appendix with a table of commonly used CNN architectures for EEG feature extraction. EEG signals are essential for various applications, including sleep monitoring, emotion recognition, motor imagery tasks, and neurofeedback therapy. They are also a key component of brain-computer interfaces (BCIs), enabling direct communication between the brain and external devices. EEG signals are non-linear and non-stationary, posing challenges in their mathematical description. Recent developments in CNNs have shown their ability to handle non-linear dependencies and decompose them into characteristic frequency components, allowing many architectures to operate on raw EEG data without additional processing. The paper discusses the challenges of EEG data processing, including noise reduction, artifact removal, and denoising. It also covers frequency and spatial components in EEG signals, as well as feature selection and extraction methods. The authors highlight the importance of data preprocessing, including filtering, averaging, and wavelet decomposition, to ensure the quality of EEG data for analysis. The paper also addresses the challenges of dataset acquisition for CNN models applied to EEG data, including the inherent noise and variability of EEG signals. It discusses transfer learning and self-supervised learning as approaches to overcome these challenges. The authors emphasize the importance of hyperparameter tuning, kernel size selection, and dimensionality reduction in CNNs for EEG analysis. The paper concludes by discussing popular CNN architectures for EEG, highlighting their ability to detect frequency patterns and their potential in various applications. The authors emphasize the importance of understanding the data processing and architecture of CNNs for fully grasping their significance in EEG feature extraction.This paper explores the use of convolutional neural networks (CNNs) for feature extraction from electroencephalogram (EEG) signals. The authors aim to provide insights into creating and fine-tuning CNNs for various applications, considering the characteristics of EEG signals and various signal processing techniques. They analyze well-known CNN architectures, categorizing them into four groups: standard implementation, recurrent convolutional, decoder architecture, and combined architecture. The paper offers a comprehensive evaluation of these architectures, covering accuracy metrics, hyperparameters, and an appendix with a table of commonly used CNN architectures for EEG feature extraction. EEG signals are essential for various applications, including sleep monitoring, emotion recognition, motor imagery tasks, and neurofeedback therapy. They are also a key component of brain-computer interfaces (BCIs), enabling direct communication between the brain and external devices. EEG signals are non-linear and non-stationary, posing challenges in their mathematical description. Recent developments in CNNs have shown their ability to handle non-linear dependencies and decompose them into characteristic frequency components, allowing many architectures to operate on raw EEG data without additional processing. The paper discusses the challenges of EEG data processing, including noise reduction, artifact removal, and denoising. It also covers frequency and spatial components in EEG signals, as well as feature selection and extraction methods. The authors highlight the importance of data preprocessing, including filtering, averaging, and wavelet decomposition, to ensure the quality of EEG data for analysis. The paper also addresses the challenges of dataset acquisition for CNN models applied to EEG data, including the inherent noise and variability of EEG signals. It discusses transfer learning and self-supervised learning as approaches to overcome these challenges. The authors emphasize the importance of hyperparameter tuning, kernel size selection, and dimensionality reduction in CNNs for EEG analysis. The paper concludes by discussing popular CNN architectures for EEG, highlighting their ability to detect frequency patterns and their potential in various applications. The authors emphasize the importance of understanding the data processing and architecture of CNNs for fully grasping their significance in EEG feature extraction.