On a $$ 4 \times 4 $$ input, applying a $$ (3, 3) $$ max-pooling layer with the default stride of $$ (3, 3) $$ produces a $$ 1 \times 1 $$ output. This output contains the maximum value of the top-left $$ 3 \times 3 $$ region, because the pooling window cannot shift further by $$ 3 $$ pixels without exceeding the input boundaries.

Google

Unlike convolutional layers—where the default stride is $$ 1 $$—deep learning frameworks set the default stride of a pooling layer equal to the pooling window size. For example, specifying a pooling window of shape $$ (3, 3) $$ automatically sets the stride to $$ (3, 3) $$. This convention ensures that consecutive pooling windows do not overlap, so each input element contributes to exactly one output element.

Default Stride Behavior in Pooling Layers

Dive into Deep Learning

Although deep learning frameworks assign default stride and padding values for pooling layers, both can be manually overridden to control the output dimensions. Padding adds extra border elements (typically zeros) to the input before pooling, allowing the pooling window to cover boundary regions. Setting the stride smaller than the window size introduces overlap between adjacent windows, whereas a larger stride increases the downsampling factor. These controls mirror the padding and stride mechanisms available in convolutional layers and are essential for achieving a desired output shape in network architectures.

Learn Before

Related