In spite of the fact that the accuracy of Motor Imagery (MI) Brain-Computer Interface (BCI) systems based on deep learning has been significantly increased in comparison with certain standard algorithms, it is still a significant challenge to comprehend the deep learning models accurately. This paper addresses the concerns by presenting a well-known deep learning model known as EEGNet, then contrasting that model with a more conventional approach known as Filter-Bank Common Spatial Pattern (FBCSP). After that, this work considers that a unique Discrete Wavelet Transform (DWT) can explain the 1-D convolution of EEGNet and that the depthwise convolution of EEGNet is comparable to the Common Spatial Pattern algorithm. In addition, the EEGNet has been made more effective by using the technique known as the Temporary Constrained Sparse Group Lasso (TSGL) developed for this study. The suggested model, a modified version of TSGL-EEGNet, was evaluated using the BCI Competition IV 2a dataset, consisting of MI tasks involving four classifications. The testing results indicate that the proposed model has achieved an average classification accuracy of 82.05 per cent on the dataset BCI Competition IV 2a. These results are higher than those achieved by TSGL-EEGNet, EEGNet, C2CM, MB3DCNN, SS-MEMDBF, using FBCSP, particularly on insensitive subjects. . Also, the testing results show 0.78 kappa value for subject dependent classification.

BCI, CNN, Motor Imagery, Temporary constrained spares group lasso