AWKNet: A Lightweight Neural Network for Motor Imagery Electroencephalogram Classification Based on Adaptive Wavelet Transform Kolmogorov-Arnold

Motor imagery electroencephalography (MI-EEG) is widely used in the neural rehabilitation field, including for hybrid device control, such as robotic arms. However, it is difficult to apply large models with good performance in consumer electronics (CE) with limited computing and memory resources. To address this challenge, this study proposes an adaptive wavelet transform Kolmogorov-Arnold network (KAN) approach named AWKNet, which uses wavelet loss to construct personalized discrete wavelet functions for MI-EEG features suited to different topics to learn an effective multiresolution wavelet transform. Second, a depth-separable convolutional layer is used to decouple the cross-channel and frequency domain features of the EEG data, and the conventional multilayer perceptron (MLP) layer is replaced based on the KAN technique. The proposed model is lightweight and improves the performance of the brain-computer interface (BCI) system. The model was employed to classify EEG signals acquired in the BCI Comparison IV 2a dataset and in a real-world environment. In both tasks, the visualization of model weights showed that the trained AWKNet consistently generates scientifically interpretable lightweight models and outperforms more advanced neural networks in terms of classification performance, which indicates that AWKNet has broader application potential in CE. All the code is deposited on GitHub (https://github.com/Songyu-EEGsignals/AdaptiveWavelets).