Transfer Learning for CNN Based Modulation Classification in Time-Variant A WGN Channels

Machine learning based modulation classifiers are often criticized for their complex training process. Although such training process can be performed offline and not impacting the actual classification speed, in time-variant channels re-training of the model is needed to optimize it for the now scenario. In this paper, a transfer learning framework is proposed to reduce the computational complexity of the re-training process. According to the simulated results, the proposed framework is able to significantly reduce the number of parameters to be re-optimized after change of noise level while maintaining a good classification performance.