We report in this paper an investigation of convolutional neural network (CNN) models for target prediction in time-locked image rapid serial visual presentation (RSVP) experiment. We investigated CNN models with 11 different designs of convolution filters in capturing spatial and temporal correlations in EEG data. We showed that for both within-subject and cross-subject predictions, the CNN models outperform the state-of-the-art algorithms: Bayesian linear discriminant analysis (BLDA) and xDAWN spatial filtering and achieved >6% improvement. Among the 11 different CNN models, the global spatial filter and our proposed region of interest (ROI) achieved best performance. We also implemented the deconvolution network to show how we can visualize from activated hidden units for target/nontarget events learned by the ROI-CNN. Our study suggests that deep learning is a powerful tool for RSVP target prediction and the proposed model is applicable for general EEG-based classifications in brain computer interaction research. The code of this project is available at https://github.com/ZijingMao/ROICNN.