To address the issue of sample imbalance in EEG obtained through rapid serial visual presentation (RSVP), a multi-task learning model for EEG classification is proposed. Firstly, a deep shared feature extraction module is established, which utilizes a convolutional neural network to automatically learn shared parameters and extract depth-related features associated with tasks. Then, a multi-task objective function is constructed based on the classification task and hyper-sphere constraint task, utilizing the joint learning of these two tasks to extract more effective discriminative features and improving the model′s generalization performance. Experiments are implemented on a public RSVP EEG dataset. Compared with commonly used EEG classification models such as DeepConvNet, EEGInception, DRL, and EEGNet, the proposed model named Multi-task EEGNet can achieve the average AUC improvement of 3. 57% , 1. 84% , 6. 22% , and 2. 09% respectively, across 32 subjects. The results indicate that the proposed multi-task learning model can extract discriminative features more fully, effectively improve model classification performance, and better solve the sample imbalance problem in EEG classification tasks.