Genetically Optimized Memristive CNNs with Online Learning for Robust EEG Emotion Recognition

This paper presents a convolutional neural network (CNN) model optimized using a genetic algorithm (GA) and an online calibration strategy for EEG-based emotion recognition. The GA is used to adjust the network structure to improve its adaptability to non-stationary physiological signals, while the calibration mechanism updates weights in real time to compensate for variations in memristor hardware. This approach improves model performance under low-precision conditions without requiring additional hardware circuits. The method is evaluated on MNIST, FashionMNIST, and SEED datasets, showing better accuracy and more stable training results compared to models without GA optimization. The overall framework supports efficient and hardware-aware learning on memristive platforms and may be applied to other edge learning tasks in similar settings.