ConvNets Pruning by Feature Maps Selection

Convolutional neural network (CNN) is one of the research focuses in machine learning in the last few years. But as the continuous development of CNN in vision and speech, the number of parameters is also increasing, too. CNN, which has millions of parameters, makes the memory of the model very large, and this impedes its widespread especially in mobile device. Based on the observation above, we not only design a CNN pruning method, where we prune unimportant feature maps, but also propose a separability values based number confirmation method which can relatively determine the appropriate pruning number. Experimental results show that, in the cifar-10 dataset, feature maps in each convolutional layer can be pruned by at least 15.6%, up to 59.7%, and the pruning process will not cause any performance loss. We also proved that the confirmation method is effective by a large number of repeated experiments which gradually prune feature maps of each convolutional layer.