Global Average Pooling is a technique used in Convolutional Neural Network (CNN) to reduce the spatial dimensions of feature maps while preserving their channel-wise information. Unlike fully-connected dense layers, which flatten feature maps into vectors, GAP uses a spatial average operation to each channel independently. Essentially, this registers the data while retaining most of the important information.
For example, if the feature map has dimensions (height, width, and channels), global average pooling turns it into a vector of size by averaging all the values in each channel. This helps simplify the model and less likely to overfit, while still being effective for tasks like classification.
The main advantage of global average pooling is that it doesnβt need extra parameters like fully-connected layers do. This makes the model smaller and easier to train and is commonly used in architectures like ResNet and GoogleNet to simplify the network and directly connect feature maps to class probabilities. For classification, the result from GAP can be passed into a softmax layer to get the probabilities for each class. Essentially, GAP replaces complex layers with a simple averaging operation, reducing model size and complexity while preserving key information from the feature maps.