Watch max, average, min, and RMS pooling slide across a pixel grid in real time. Understand how CNNs reduce spatial dimensions while preserving the features that matter.
Sign in to save progress
Max Pooling: Takes the strongest signal
Picks the largest value in each window. Great for preserving the most prominent feature, such as the brightest edge or strongest activation.
max([120, 80, 60, 240]) = 240
Same 8×8 input, same 2×2 kernel, stride 2 → 4×4 output. Max Pooling retains sharp features (higher peak values); Average Pooling is smoother (values blend together).
Preserves strongest activations. Sharper contrast.
Blends values together. Smoother, lower peaks.
Max Pooling range
Average Pooling range
Spatial Invariance
Small shifts or translations of an object in the input don't significantly change the pooled output. This helps CNNs recognise the same feature regardless of its exact position.
Dimensionality Reduction
Pooling shrinks the spatial size of feature maps, reducing the number of parameters and computation in subsequent layers. A 2x2 max pool with stride 2 cuts width and height by half.
Noise Suppression
Max pooling naturally ignores low activations by taking only the peak value in each region. This helps the network focus on strong, meaningful signals and ignore background noise.
In modern CNNs like ResNet, Global Average Pooling collapses each feature map to a single value per channel, replacing the huge flattened fully-connected stacks of older architectures like VGG; only one small final classifier layer remains. This dramatically reduces parameters and overfitting.