Average Pooling

Average pooling is a way for a neural network to shrink a feature map while retaining the general strength of nearby signals. Rather than asking “what was the strongest feature here?” it asks “how present was this feature across this small region?” This makes it a calm, smoothing form of spatial compression.

How it works
In a convolutional network, a feature map is a grid of numbers produced by a learned filter. An average-pooling layer slides a fixed window, such as 2×2, across that grid and replaces each window with its arithmetic mean. With a 2×2 window and stride 2, a 4×4 feature map becomes 2×2: each output value summarizes four input values. Pooling has no learned weights; its window size, stride, and padding are chosen when the model is designed.

What it preserves—and loses
Average pooling reduces spatial resolution, computation in later layers, and memory use. It also makes a feature less sensitive to small shifts: moving an activation by one pixel changes a local average less abruptly than changing one exact location. The trade-off is detail. A sharp, highly meaningful activation can be diluted by surrounding low values, whereas max pooling preserves the strongest response.

• 2D average pooling downsamples local regions inside a network.
• Global average pooling averages each entire feature map into one number, producing one feature per channel.

Training and design implications
During backpropagation, the gradient arriving at one pooled output is distributed equally across the input values in its window. This creates smooth, stable gradient flow, but it does not teach the layer which location mattered most. In PyTorch, torch.nn.AvgPool2d performs local pooling, while AdaptiveAvgPool2d((1,1)) implements global pooling regardless of input size. Use it when broad feature presence matters more than precise location; avoid aggressive pooling early in a network when fine spatial structure is still needed.