If you are building an image classification network and want to reduce the spatial size of the feature maps to lower computational complexity while retaining the most prominent features, which layer type should you insert after a convolutional layer?

In Convolutional Neural Networks (CNNs) designed for tasks like image classification, managing the spatial dimensions of feature maps is critical to prevent overfitting and control computational overhead. The standard architectural approach to address this is the insertion of a pooling layer (most commonly Max Pooling or Average Pooling) after a convolutional layer. Pooling operations downsample the feature maps by applying a sliding window (e.g., 2x2 with a stride of 2) across the width and height of the input. Max pooling extracts the maximum value within each window, which represents the most prominent feature activation in that local region. This reduces the total number of parameters and computation in the subsequent layers of the network. Furthermore, pooling provides translation invariance, meaning the network can still recognize a feature even if its position in the image shifts slightly. Let's evaluate the distractors: Option A is incorrect because adding another convolutional layer does not inherently downsample the features unless configured with a high stride, and typically increases computational complexity. Option B describes a dense layer, which is used at the end of the network for classification and would catastrophically increase the parameter count if connected directly to a high-resolution feature map. Option C describes a dropout layer, which is a regularization technique that randomly deactivates neurons during training to prevent overfitting, but does not downsample or reduce the spatial dimensions of feature maps. For the exam, associate pooling layers with dimensionality reduction, downsampling, and spatial feature retention.