Dilated convolution

As explained in Sutskever's List

A dilated convolution applies a convolutional filter with spaced sampling positions to enlarge the receptive field without proportionally increasing filter size.

Dense prediction needs broad image context without losing the fine details required for pixel-level labels. Dilated convolution spaces out a filter’s sampling positions, enlarging its receptive field without enlarging the filter or reducing resolution through pooling. Stacking dilation rates of 1, 2, 4, and 8 expands the field of view while keeping each layer’s parameter count unchanged. section 3.5

Dense prediction with VGG-16

Yu and Koltun’s 2015 approach removed pooling and striding layers from VGG-16 and replaced them with dilated convolutions, preserving high-resolution feature maps while gathering broader context. section 3.5

Figure 3.6 Dilated convolutions rapidly expand the receptive field without increasing computational cost. Each dot represents a convolutional kernel element. As dilation increases (from left to right), the receptive field expands exponentially, capturing broader spatial context while keeping the parameter count the same. Used with permission granted by the lead author (Fisher Yu).

Compare how increasing dilation spreads a filter’s sampling positions and expands its receptive field without changing its parameter count. section 3.5

How Sutskever's List covers it

How deep it goes

Explains spaced sampling and stacked dilation rates, comparing receptive-field growth with parameter growth.

Examples

Uses dilation rates 1, 2, 4, and 8 and a modified VGG-16 backbone.

What you can do after

Evaluate ways to expand receptive fields while retaining resolution and avoiding larger kernels or pooling.

How it connects

Connects receptive fields, pooling, semantic segmentation, and context modules.

liveBookmanning.com

Read Sutskever's List on liveBook Get the book at manning.com

Related concepts

• stridecontrasts with · A stride is the step interval between successive positions or elements accessed, selected, or processed along a dimension or sequence.
• convolutionkind of · Convolution is an operation that applies a kernel across an input, combining values from overlapping local regions with shared weights to produce outputs.
• receptive fieldproduces · A receptive field is the set or region of input elements whose information can influence a particular neural-network unit, representation, or output.
• convolution kerneluses · A convolution kernel is a small array or tensor of weights applied across local input regions to compute weighted responses and detect or transform patterns.

This explainer was generated by Manning's knowledge graph from the text of Sutskever's List. The section links point to the parts of the book it draws on.