Feature Maps

Short Definition

Feature maps are the intermediate representations produced by neural network layers that encode detected patterns or features across the input.

Definition

A feature map is an output tensor generated by applying a transformation—most commonly a convolution followed by a nonlinearity—to an input. Each feature map represents the presence and spatial distribution of a learned feature, such as edges, textures, or higher-level concepts.

Feature maps are learned descriptions of data.

Why It Matters

Feature maps are the primary carriers of information inside convolutional networks. Their structure, resolution, and depth determine what patterns the network can represent, how information flows across layers, and how effectively the model can generalize.

What the model sees is what it can learn.

How Feature Maps Are Created

Feature maps typically arise from:

convolution operations with learned filters
activation functions (e.g., ReLU)
normalization layers
pooling or strided operations

Each filter produces one feature map.

Minimal Conceptual Illustration

Input → Convolution + Activation → Feature Maps

Spatial Structure

Feature maps preserve spatial relationships:

each location corresponds to a region in the input
spatial arrangement reflects where features occur
resolution decreases as depth increases (in many architectures)

Space encodes meaning.

Channels and Depth

The number of feature maps equals the number of output channels in a layer.

shallow layers: few channels, high resolution
deep layers: many channels, low resolution

Depth trades space for semantics.

Feature Hierarchies

Across layers, feature maps evolve:

early layers: edges, corners, simple textures
mid layers: motifs, parts, repeated structures
deep layers: objects, regions, abstract concepts

Abstraction emerges hierarchically.

Feature Maps and Receptive Fields

Each element in a feature map corresponds to a receptive field in the input. As depth increases, receptive fields grow, allowing deeper feature maps to encode broader context.

Context expands with depth.

Feature Maps vs Activations

Feature map: the structured tensor output of a layer
Activation: the value of a unit within a feature map

Maps organize activations.

Feature Maps and Interpretability

Visualizing feature maps can:

reveal learned patterns
diagnose training issues
expose overfitting or bias
aid model debugging

Visibility aids understanding—but is incomplete.

Feature Maps in Modern Architectures

Feature maps are used in:

CNNs and ResNets
hybrid CNN–Transformer models
attention mechanisms (as value/key tensors)
dense prediction models

Representation is universal.

Limitations

Feature maps:

are not inherently interpretable
can encode spurious correlations
may collapse under excessive downsampling
are sensitive to distribution shift

Seeing features is not seeing intent.

Common Pitfalls

assuming high activation implies importance
excessive channel growth without benefit
aggressive early downsampling
ignoring calibration and robustness of features
conflating visualization with explanation

Features must be validated.

Summary Characteristics

Aspect	Feature Maps
Role	Intermediate representation
Structure	Spatial + channel-wise
Evolves with	Depth
Interpretability	Partial
Sensitivity	Task-dependent

Neural Network Lexicon