What is the difference between max pooling and average pooling over a CNN feature map?

## What pooling is A pooling layer slides a window (commonly 2x2 or 3x3) over a feature map with a stride (commonly 2) and replaces each window with a single summary number. Three properties matter and are frequently misstated in interviews: 1. **It has no learnable parameters.** The reduction rule is fixed in advance. Nothing about pooling is trained. 2. **It acts per channel.** A map with C channels goes in and C channels come out. Pooling never mixes or reduces channels; it only reduces spatial extent. 3. **It changes the spatial shape by the usual arithmetic.** A 2x2 window with stride 2 on a 32x32xC map yields 16x16xC. A 3x3 window with stride 2 (an overlapping pool) yields overlapping windows but the same halving behaviour up to edge handling. ## The two reductions **Max pooling** outputs the maximum activation in the window. Because a convolutional filter's activation is "how much did this pattern match here", the maximum answers a detection question: *did this pattern appear anywhere in this neighbourhood?* The exact position inside the window is thrown away, which gives a small amount of tolerance to where the evidence sits, and the magnitude of the strongest evidence is carried forward undamaged. **Average pooling** outputs the mean of the window. It answers a density question: *how strongly, on average, did this pattern match around here?* It keeps distributed, low-contrast structure and suppresses single-pixel excursions, but it necessarily attenuates a lone strong response: one value of 8.0 in a 2x2 window of zeros comes out as 2.0 under averaging and as 8.0 under max. ## Choosing between them by what the signal looks like The honest rule is: pick the statistic that matches how the evidence is distributed in your data. - **Sparse, high-contrast evidence favours max.** Consider a pathology or mammography tile where the class-defining feature is a bright three-pixel calcification against flat tissue. Under 2x2 average pooling that value is divided by four at every pooling stage; after three stages a factor of 64 separates it from where it started, and it can fall below the background variation that the rest of the tile contributes. Max pooling propagates it intact. - **Distributed, textural evidence favours averaging.** Consider grading fabric by weave regularity, where no single pixel is decisive and the class lives in the average roughness of a region. Here max pooling reports the most extreme pixel of every window, which is essentially the noise ceiling, and it will report a similar number for a good and a bad sample. The mean is the discriminative statistic. - **Noise behaviour follows from this.** Max pooling is sensitive to outliers and to sensor spikes, because an artefact that is brighter than the real signal wins the window. Average pooling is robust to spikes but can be dominated by a bright background. ## Why pool at all Three reasons, all worth stating: - **Compute and memory.** Halving both spatial dimensions quarters the number of positions every later layer must process, which is what makes deep stacks affordable. - **View of the input.** After a stride-2 reduction, one step in the output corresponds to two pixels of input, so every subsequent layer's window covers twice as much of the original image. This is the mechanism by which a network's field of view grows quickly with depth rather than a couple of pixels at a time. - **Small-shift tolerance.** Summarising a window means the exact placement of a feature inside that window stops mattering, which is useful when you care that a pattern is present, not where within a two-pixel neighbourhood it sat. ## What it costs Pooling destroys spatial precision, and that precision is unrecoverable. For a classification head this is acceptable and often helpful. For a task that must report *where* something is at pixel accuracy, aggressive early pooling is a design decision that has to be paid back later through architectural choices that restore resolution. The other cost is that the rule is fixed: max pooling asserts, as a prior baked into the architecture, that the maximum is the useful summary. When it is not, the network cannot override it. ## Typical usage In classical convolutional trunks, max pooling (or a strided reduction) is used between blocks, and averaging appears at the very end, over the whole map, to produce one number per channel for the classifier. A common interview trap is asserting that pooling reduces the number of feature maps, or that it has trainable weights; both are wrong, and both are easy to check against the shape arithmetic.