A class of Electroencephalogram (EEG) generated when a person is present with a stimuli are Error-related potentials (ErrP). Classification of ErrP revolutionized the domain of Brain-Computer Interface (BCI), however, due to its poor classification accuracy, it is difficult to use it in practical applications. The use of deep learning techniques such as Convolutional neural networks (ConvNets) has been growing because of its end-to-end learning capabilities and classification performance in other fields such as computer vision, speech processing and text synthesis. Very little work has been published towards classification of EEG or its classes using deep learning approaches because of its unexplored benefits in BCI applications. In this paper we propose a novel deep ConvNet architecture to accurately classify Error-related potentials. Using the very recent advances from the field of machine learning, such as batch normalization and dropout layers, we designed a new deep ConvNet architecture. We evaluate and compare the performance of two proposed deep ConvNets by finding its classification accuracy for EEG recordings from same and cross sessions as well as cross subjects for different trials of individual subjects. We see that the net proposed later with newer layers performs very well by improving the mean performance by 4%.