Textbook of AI

A Feature Map (sometimes called an activation map) is the output of a convolutional layer, typically arranged as a 3D tensor with spatial dimensions (height and width) and a channel dimension where each channel corresponds to a different learned feature. When a convolutional kernel slides over its input and computes dot products with the local patches, the resulting 2D response grid is the feature map for that kernel. A layer with $K$ kernels produces $K$ stacked feature maps, forming the output tensor.

In early layers of a CNN trained on images, feature maps typically correspond to low-level visual features: edges at various orientations, colour blobs, textures. In middle layers, feature maps represent combinations of edges forming shapes, textures, and simple parts. In deeper layers, feature maps encode high-level semantic concepts: object parts, specific objects, and abstract attributes. The progressive abstraction of feature maps from pixels to concepts is the defining characteristic of hierarchical representation learning in CNNs.

Feature maps are the input to subsequent layers and, in classification networks, are eventually reduced (via pooling, global average pooling, or flattening) into a vector that feeds a final classifier. Feature maps are also the basis of many visualisation and interpretation techniques: class activation mapping (CAM) and Grad-CAM highlight regions of feature maps most responsible for a prediction, producing interpretable heatmaps. Skip connections in architectures like U-Net and ResNet pass feature maps directly from earlier layers to later ones, preserving spatial detail and improving gradient flow.