Neural Network Lexicon

Short Definition

Dilated convolutions expand a convolution’s receptive field by inserting gaps between kernel elements without increasing parameter count or reducing resolution.

Definition

Dilated convolutions (also called atrous convolutions) modify the standard convolution operation by spacing kernel elements apart according to a dilation rate. This allows the network to aggregate information from a wider area of the input while preserving spatial resolution and using the same number of parameters.

Dilations see farther without pooling.

Why It Matters

Standard convolutions grow receptive fields slowly and pooling sacrifices spatial detail. Dilated convolutions provide a third option: large receptive fields with dense output maps. This is especially valuable in tasks requiring global context and precise localization.

Context without compression.

Dilation Rate

The dilation rate determines the spacing between kernel elements:

• dilation = 1: standard convolution
• dilation = 2: one gap between kernel elements
• dilation > 2: increasingly sparse sampling

Dilation controls reach.

Minimal Conceptual Illustration

Standard kernel: X X X
Dilated kernel: X . X . X (dilation = 2)

Effect on Receptive Fields

Dilated convolutions increase the receptive field exponentially with depth while keeping feature map resolution constant.

Receptive field grows without downsampling.

Parameter Efficiency

Dilated convolutions:

• do not add parameters
• reuse the same kernel weights
• increase context at no parameter cost

Efficiency comes from structure.

Relationship to Stride and Pooling

Operation	Resolution	Receptive Field	Information Loss
Pooling	Reduced	Increased	Yes
Stride > 1	Reduced	Increased	Yes
Dilation	Preserved	Increased	No

Dilations preserve detail.

Common Use Cases

Dilated convolutions are commonly used in:

• semantic segmentation
• dense prediction tasks
• audio and time-series modeling
• wave-based architectures
• hybrid CNN–attention models

Dense outputs need dense context.

Gridding Artifacts

A known issue with large dilation rates is gridding artifacts, where sparse sampling causes checkerboard or blind-spot patterns.

Wide reach can miss details.

Mitigation Strategies

To reduce gridding effects:

• combine multiple dilation rates
• use dilation pyramids
• interleave standard and dilated convolutions
• apply multi-scale feature aggregation

Context must be balanced.

Dilated Convolutions vs Attention

Dilated convolutions:

• encode structured, local-to-global bias
• are efficient and deterministic

Attention:

• models global interactions explicitly
• is more flexible but computationally heavier

Dilations are structured context; attention is adaptive context.

Limitations

Dilated convolutions may:

• struggle with irregular global dependencies
• introduce aliasing artifacts
• require careful architectural tuning
• underperform when global reasoning dominates

Bias must match the task.

Common Pitfalls

• using large dilation rates too early
• stacking dilations without multi-scale design
• assuming dilations replace attention entirely
• ignoring gridding artifacts
• misaligning dilation with task resolution needs

Reach without control harms learning.

Summary Characteristics

Aspect	Dilated Convolutions
Receptive field growth	Large
Resolution	Preserved
Parameter cost	None
Information loss	None
Risk	Gridding artifacts

Related Concepts

• Architecture & Representation
• Convolution Operation
• Receptive Fields
• Stride and Padding
• Pooling Layers
• Feature Maps
• Semantic Segmentation
• Attention Mechanisms