Global Entropy Pooling layer for Convolutional Neural Networks

Katarzyna Filus, Joanna Domańska

Abstract: We propose a novel Global Entropy Pooling (GEP) layer for Convolutional Neural Networks (CNNs). This is the first approach that uses the Entropy value directly for pooling rather than creating a weighting mechanism for feature maps obtained via convolution. This way, we use the ”amount” of information (expressed by Entropy) to describe the feature maps in contrast to the most commonly used averaging (Global Average Pooling, GAP). We use our method in a rather unusual way. We replace GAP with GEP and use the same feature extractor and classifier weights for the prediction — with no additional training. Our layer can be easily integrated (instead of the GAP layer) into various pre-trained CNN models. From a technical perspective, it is possible to integrate GEP without additional training, because it does not contain any parameters that have to be trained. Also, its input and output have the same format as the GAP’s one. We examine how the replacement impacts the model behavior. Despite a completely different form of pooling, utilization of the GEP layer in the examined pre-trained models (InceptionV3, Xception and InceptionResNetV2) allowed us to obtain some accuracy improvements in comparison to the original networks on the subsets of the ImageNet dataset. Moreover, networks modified with our layer exhibit an interesting property — they are more ”confident” in the predictions they return. It manifests itself through higher values of the probabilities assigned to correctly predicted classes. This opens a possibility to use not only the initial layers of the pretrained network, but also to modify network’s middle parts and reuse its classifier trained on an original model as well.