Max pooling versus strided convolution

WHAT THIS TESTS Whether you understand why CNNs downsample and can reason about two competing ways to do it rather than reciting one.

A GOOD ANSWER COVERS A pooling layer reduces the spatial resolution of feature maps. Max pooling slides a window, commonly two by two with stride two, and outputs the maximum value in each window, halving height and width. Its purposes are reducing computation and memory downstream, enlarging the effective receptive field of later layers, and granting a small amount of translation invariance because the max is unchanged if the activating feature shifts slightly within the window. Pooling has no learnable parameters and is applied per channel. The alternative, strided convolution, achieves the same spatial reduction but the downsampling is learned jointly with feature extraction, so the network can choose what to keep rather than blindly taking the max.

COMMON WRONG ANSWERS Saying pooling has weights to train; it does not. Claiming pooling always improves accuracy; many modern architectures replace it with strided convolutions. Confusing max pooling with average pooling without noting average pooling smooths rather than selects the strongest response.

LIKELY FOLLOW-UPS When would you prefer average pooling. Why have some architectures dropped pooling entirely. How does pooling affect backpropagation, given the gradient only flows to the max element.

ONE CONCRETE EXAMPLE A feature map of size four by four with a two by two max pool, stride two, becomes two by two, each output being the largest of four inputs. If a detected edge shifts by one pixel but stays within the same window, the pooled output is identical, illustrating local invariance. A strided three by three convolution over the same input also produces a smaller map but learns weights, potentially retaining gradient information that max pooling would throw away.