Recently, deep learning methods received great attention in pattern recognition and especially classification problems in medicine fields. Deep learning methods are branch of machine learning which use neural networks with deep structure, more and specific layers and more parameters to extract appropriate features to solve pattern recognition problems better and more precise. Also, atrial fibrillation (AF) is a common arrhythmia which has been investigated in several studies. In this paper, we used a division of deep learning method called stacked autoencoder to classify atrial fibrillation (AF). Autoencoder networks, like traditional neural networks use a softmax layer in the last layer and always seeking the optimization problem in training parameters for classification. A basic difference between AE and traditional networks is the concept of deepening these networks, setting some autoencoders in stack form, i.e. the second autoencoder analyzes output of the first autoencoder, the third autoencoder analyzes the output of the second autoencoder and so on. We used electrocardiogram (ECG) signals from MIT-BIH database and computed commonly spectral, time and non-linear features from them. First extracted features were evaluated using statistical test, analysis of variance (ANOVA) and selected significant features then used stacked autoencoders as parallel form to primary classify AF and normal samples. Then, final decision performed using the ensemble averaging method. The average accuracy for classifying AF and normal samples achieved 95.5%. This results showed that the proposed method was successful and can be used for other classification problems.