Automatic modulation recognition (AMR) can automatically estimate the modulation type under the condition that the signal is unknown at all. A deep learning based AMR is proposed. The proposed method can update the filter taps through waveform optimization, which can filter the signal samples in order to overcome the unfavorable effects of the transmission channels. In the proposed method, a feedback path exists between the recognition network and the inverse-channel filter. According to the experiments from an open-source dataset, the proposed feedback-structured method can increase the recognition rate compared with the traditional deep learning methods. Specially, compared with the CNN based method, the recognition rate has increased by about 7%.