We present an efficient and robust computational model for brain state interpretation from EEG single trials. This includes identification of the most relevant time points and electrodes that may be active and contribute to differentiation between the mental states investigated during the experiment. The model includes a regularized logistic regression classifier trained with cross-validation to find the optimal model and its regularization parameter. The proposed framework is generic and can be applied to different classification tasks. In this study we applied it to a classical visual task of distinction between faces and houses. The results show that the obtained single trial prediction is significantly better than chance. Moreover, correct choice of the regularization parameter significantly improves classification results. In addition, the obtained spatial-temporal information of brain activity can give an indication to correlated activity of regions of the brain (spatial) as well as temporal activity correlations between and within EEG electrodes. This spatial-temporal analysis can render a far more holistic interpretability for visual perception mechanism without any a priori bias on certain time periods or scalp locations.