Dilated Convolution

A dilated convolution lets a small filter see a wider area without using more weights. Instead of sampling every neighboring position, the filter spaces its samples apart by a chosen dilation rate. This gives the network more context while keeping the feature map at the same resolution.

Mechanically, dilation inserts gaps between the positions touched by a convolutional kernel. A three-by-three filter with a larger dilation can cover the span of a much bigger filter, but still learns only nine weights. Stacking layers with different dilation rates grows the receptive field quickly, which is useful for dense prediction, audio modeling, and temporal sequence processing.

The advantage is efficient context: models can see larger structures without repeated pooling or expensive large kernels. The tradeoff is that sparse sampling can create gridding artifacts, miss fine local detail, or interact poorly with hardware kernels optimized for ordinary convolutions. It usually works best when mixed with standard convolutions, skip connections, or multi-scale modules.

Open questions concern how much dilation should be fixed by design rather than learned from data. Adaptive or deformable variants can choose more flexible sampling patterns, but add complexity and instability. Attention and state-space models now compete for many long-context roles, so the enduring niche of dilation depends on where convolutional locality remains the most efficient bias.