What is the role of the pooling layer in a CNN?

Instruction: Explain what pooling layers do and why they are important.

Context: This question tests the candidate's knowledge of the architecture of convolutional neural networks, focusing on the specific role of pooling layers.

Updated

Example Answer

Pooling aggregates nearby activations. With a stride greater than one, it usually reduces spatial resolution and computation in later layers. Max pooling keeps the largest value in each window; average pooling summarizes the window. It can add tolerance to small shifts, but it does not guarantee translation invariance or know which pixels are useful foreground. For a small-object detector, aggressive pooling might erase the very detail the task needs, so I would measure that trade-off.

Make it your own

Illustrative test: shift an image by one pixel and compare logits and small-object recall before and after changing pooling or downsampling.

Why this works

The test connects an architectural choice to observable behavior. Aliasing and lost spatial detail can offset compute savings. Follow-up: how could anti-aliased downsampling change shift sensitivity?

Sources: Making Convolutional Networks Shift-Invariant Again.

Related Questions