As part of the PhysioNet/Computing in Cardiology Challenge 2017, this work focuses on the classification of a single channel short electrocardiogram (ECG) signal into normal, atrial fibrillation (AF), others and noise classes. To this end, we propose a shallow convolutional neural network architecture which learns suitable features pertaining to each class while eliminating the need to extract the traditionally used ad hoc features. In particular, we first developed a robust R-peak detector and stacked sequence of fixed number of detected beats with R-peaks aligned. These stack of beats corresponding to a segment of ECG record are classified into one of the four aforementioned classes. To improve the robustness, multiple classifiers were trained to classify these segments. Overall record classification was then generated using an voting scheme from the classification results of individual segments. Our best submission result during the official phase has a score of 71% with F1 scores of 86%, 73% and 56% respectively for normal, AF and other classes respectively. © 2017 IEEE Computer Society. All rights reserved.