Training Neural Networks Using Features Replay

Finally, we apply our method to training deep convolutional neural networks, and the experimental results show that the proposed method achievesfasterconvergence,lowermemoryconsumption,andbettergeneralization errorthancomparedmethods.