Feature map
A feature map is the output produced by applying a single convolutional filter to its input. Each value in the feature map records how strongly the pattern...
What a Feature Map Represents
A feature map is the output produced by applying a single convolutional filter to its input. Each value in the feature map records how strongly the pattern encoded by that filter appears at the corresponding spatial location in the input. If the filter has learned to detect vertical edges, the feature map will have large positive values wherever vertical edges exist in the input and small values elsewhere.
In a typical convolutional layer, many filters are applied simultaneously, each producing its own feature map. These are stacked along the channel dimension to form a three-dimensional tensor: height by width by number of filters. The depth of this tensor, the number of feature maps, represents the variety of distinct patterns the layer can detect.
Exam tip: Distinguish between feature map and activation map. These terms are used interchangeably in most contexts, but some authors use activation map specifically to refer to the feature map after a nonlinear activation function has been applied, while feature map refers to the raw linear output of the convolution.
Hierarchy of Feature Maps Across Layers
The feature maps produced by shallow layers of a convolutional network tend to represent simple patterns: edges, blobs, color gradients, and textures. As depth increases and each layer builds on the feature maps of the previous one, the patterns represented become progressively more complex and semantically meaningful. Intermediate layers may encode parts of objects such as eyes or wheels, while the deepest layers before the classification head encode object-level concepts.
This hierarchical structure was confirmed visually in early deep learning research by deconvolution-based visualization methods that revealed what input patterns maximally activate specific feature map channels. The finding aligned with theories about how the biological visual cortex processes information in stages from simple to complex.
Understanding this hierarchy is practically useful when deciding where to insert a new task-specific head into a pretrained backbone. Features from shallow layers contain more spatial detail and are appropriate for tasks requiring precise localization, while deep features are more useful for global recognition tasks.
Channel Depth, Computational Cost, and Design Choices
The number of feature maps in a layer is a design hyperparameter with direct implications for model capacity and computational cost. Doubling the number of filters doubles both the number of parameters in that layer and the computational cost of computing the layer's output. The classic doubling-with-depth pattern used in VGG, ResNet, and similar architectures reflects a conscious trade-off: as spatial resolution decreases with depth, channel count increases to compensate, keeping the total information content roughly constant across stages.
Depthwise separable convolutions, introduced in architectures like MobileNet, decouple spatial filtering from channel mixing. A depthwise convolution applies a single filter per input channel, and a pointwise 1-by-1 convolution then mixes channels. This produces similar feature maps to a standard convolution at a fraction of the computational cost, enabling effective models on mobile and edge devices.
In an interview, being able to estimate the number of parameters and multiply-accumulate operations for a convolutional layer from its filter dimensions and channel counts demonstrates quantitative fluency with neural network design.