3.4.1. Convolutional Neural Networks

Convolutional Neural Networks (CNNs) are designed to process grid-like data, especially in image analysis. They use a layered approach, where each layer applies filters to small sections of data and extracts specific features. A major strength of CNNs is their ability to automatically learn and identify complex spatial relationships in data. However, they require extensive data and computing resources to avoid overfitting and achieve better results.

CNNs are now also widely used in other domains such as emotion recognition. EEG (brain waves) or ECG (heart activity) are usually recorded as structured, grid-like data over time, where each “grid” containing specific values from different sensors or electrodes. CNNs apply filters, step by step, that capture key features, such as frequency changes or shifts in signal intensity, which may reveal emotional states like stress or calm. By learning these patterns, CNNs become able to tell emotions apart based on the distinctive features in each signal.

Rahim et al. used AlexNet-based CNNs on ECG and GSR signals for emotion recognition, achieving high accuracy, notably 93% when combining both signals on the AMIGOS dataset. When working with EEG signals specifically, CNN architectures were modified to detect both spatial and temporal patterns. For example, Tang and He introduced depthwise and separable convolutions in CNNs to capture temporal, person-related, and event-related features, achieving state-of-the-art results on DEAP and DREAMER datasets, while Zhang et al. introduced SparseDGCNN, a CNN model with a sparse graph structure to focus on key EEG features. Other studies also explored efficient, hardware-friendly CNN models, such as BioCNN by Gonzalez et al. , which was designed for wearable devices to perform in real-time.