Historically, most of the channel encoding and decoding algorithms have been designed to deal with and evaluated under the additive white Gaussian noise (AWGN) channel. However, in the reality, the channel is far more complicated than the AWGN assumption. Traditionally, qualizers are employed to combat the channel effects and frequency-selective fading of wireless channels. In this article, we take the advantage of deep learning approaches to handle the various channel distortions, by proposing an end-to-end approach. To train the model efficiently, the training data is obtained by simulation where the encoding process and the channel effects are viewed as a complete black box. This method can also be applied to time- varying channels for simultaneously channel estimation and symbol detection. Simulation results show that the deep learning based decoders have the ability to learn the complicated encoder function and address various channel effects. Furthermore, the deep learning based method provides an end-to-end approach, leading to a better performance in the channels with various distortions. This article represents the first step for a universal framework of information recovering from a large range of channel codes and channels.