One third of all epileptic patients is resistant to medical treatment. The construction of machines, that would detect an imminent epileptic attack based on EEG signals, represents an efficient alternative, that would help to increase their quality of life. In this article we described the implementation of an automatic detection method, based on the signal of different frequency sub-bands, using topographic maps and deep learning techniques. We constructed an ensemble of five convolutional neural networks, to classify samples of each sub-band and chose the final decision by a majority voting. The ensemble obtained 99.20% accuracy, 96.48% sensitivity and 99.27% specificity when detecting seizures of one patient. Moreover, when the networks were trained with samples taken randomly from the inter-ictal intervals, we identified on 18 of 21 seizures some false positive classifications close to the seizure onset, thus anticipating the detection of the seizure. Such misclassifications did not occur when training was performed with samples taken within five minutes of the seizure onset.