The problem of extracting a single trial visual evoked potential signal, buried in the ongoing EEG activity and measurement noise has been investigated. A method for detecting the stimulus related part of the brain activity resulting from visual flash stimulation is presented. A mixed approach, based on neural networks, non-linear auto regressive moving average (NARMA) modeling which combines gradient radial basis functions (GRBF) and orthogonal forward regressions (OFR) is used. The hidden node at each GRBF node detects and reacts to the gradient of the observed data in order to counter the level and trend of time series. In this way, non-stationary and non-linear nature of the problem is accounted and the proposed neural network's predictive ability is improved.