Receptive Field

A convolutional network does not inspect an entire input at once. Each unit in a feature map “sees” a particular region of the original input, and that region is its receptive field. As layers stack, these local views combine into progressively wider context.

How it grows through a network
In a first convolutional layer with a 3×3 kernel, each output value depends on a 3×3 patch of input values. A second 3×3 convolution builds on nearby first-layer outputs, so one of its values depends on a 5×5 region of the original input—not merely another 3×3 patch. Kernel size, stride, pooling, and dilation determine how quickly the receptive field expands:

• Larger kernels directly cover more input positions.
• Stride and pooling enlarge the input area represented by later features, while reducing spatial detail.
• Dilated convolutions spread kernel positions apart, gaining wider context without adding many parameters.
• Padding affects which input regions can influence features near an edge.

Why its size matters
A unit needs enough context to recognize the pattern its layer is responsible for. Early layers with small receptive fields can detect local edges or textures; deeper units need a wider field to relate distant parts. If the field is too small, a network cannot use the context required for its task. Enlarging it carelessly, though, can discard fine detail through aggressive striding or pooling and increase computation with very large kernels. Architectures such as ResNet commonly enlarge context gradually through stacked 3×3 convolutions, while segmentation-style networks combine deep, broad-context features with earlier high-resolution ones.

Theoretical versus effective context
The calculated, or theoretical receptive field, identifies every input location that could affect an output. During training, influence is rarely uniform: central locations usually contribute much more strongly than distant locations. This is the effective receptive field. It helps explain why a model can have a mathematically large receptive field yet still behave as though it relies mostly on nearby evidence. Receptive-field design is therefore a balance between context, spatial precision, parameter cost, and the paths through which gradients can learn useful dependencies.