EEG event-related potentials, and the P300 signal in particular, are promising modalities for brain-computer interfaces (BCI). But the nonstationarity of EEG signals and their differences across individuals have made it difficult to implement classifiers that can determine user intent without having to be retrained or calibrated for each new user and sometimes even each session. This is a major impediment to the development of consumer BCI. Recently, the EEG BCI literature has begun to apply convolutional neural networks (CNNs) for classification, but experiments have largely been limited to training and testing on single subjects. In this paper, we report a study in which EEG data were recorded from 66 subjects in a visual oddball task in virtual reality. Using wide residual networks (WideResNets), we obtain state-of-the-art performance on a test set composed of data from all 66 subjects together. Additionally , a minimal preprocessing stream to convert EEG data into square images for CNN input while adding regularization is presented and shown to be viable. This study also provides some guidance on network architecture parameters based on experiments with different models. Our results show that it may possible with enough data to train a classifier for EEG-based BCIs that can generalize across individuals without the need for individual training or calibration.