Deep Convolutional Neural Networks: Local Features vs Global Features In the paper Learn To Pay Attention an attention mechanism is developed which takes into account both local and global features.
It is written that the local features are basically the feature maps extracted from the intermediate convolutional layers of the network, while the global features are the feature maps output from the whole CNN architecture. So the global features are typically given as input to fully-connected layers.
But conceptually why is this true? I can't fully understand local features and global features extracted in a CNN
 A: Feature maps contain local features because, if the receptive field is smaller than the size of the input image, then the feature at the top left of the map will not depend at all on the value of the pixels in the bottom right of the input image -- you could change those pixel values to whatever you want without changing the features in the top left. Even when the receptive field is not strictly smaller than the image, it's an empirical fact that pixels "far away" from their corresponding feature don't have much of an affect on the value of that feature. Therefore these features are called "local", because they only depend on the pixels local to them.
Note: "far away" and "local" imply that there's some distance function between pixels and features, and the way this works is you could imagine scaling and overlaying the feature map on the image such that the top left / bottom right feature is on top of the top left / bottom right pixel of the image.
Global features are global because they depend on all pixel values of the image. The fully connected layer producing the global features takes as input all the conv features of the last convolutional layer, so this is true.
