Pooling layers
Presence over position
Your feature maps now say "vertical edge at pixel (3,7), strength 3". Ask yourself: for recognizing a cat, does the (3,7) part matter? If the whole cat shifted two pixels left, it would still be exactly as much a cat. After detection, the strength of the evidence matters far more than its precise pixel address, and carrying full-resolution maps through the whole network is paying premium price for information you barely want.
Max pooling is the standard deal: slide a small window (classically 2×2) over the feature map and keep only the STRONGEST value in each window. The map halves in width and height, 75% of the numbers gone, but every strong detection survives. You keep "definitely present, roughly here" and drop "exactly which pixel".
The deal pays three ways:
- Compute: every following layer works on a quarter of the numbers.
- Growing receptive field: after pooling, one cell covers a wider patch of the original image, so the next filters see bigger structures without bigger kernels.
- Shift tolerance: a small nudge of the input barely changes the pooled map, as you are about to verify yourself.
Run a max-pool by hand
A 4×4 feature map with two strong detections (9 and 8). Pool it with 2×2 windows, each window keeps only its champion. Then nudge the whole pattern one pixel and see how much the output actually cares.
The pyramid this builds
Now zoom out and watch what alternating convolution and pooling does across a whole network. Each conv layer adds MORE maps (more questions asked); each pooling layer SHRINKS them (less position kept). A typical journey: 224×224×3 pixels in → 112×112×64 → 56×56×128 → … → 7×7×512 at the top.
Read that trend for what it is: the network continuously trades WHERE for WHAT. Early on: precise positions, simple features. At the top: barely any geography left, but hundreds of channels of rich evidence, "wheel-ness present", "fur-ness absent". Exactly what a final classifier wants to consume.
Two footnotes for completeness: average pooling (take the mean instead of the max, gentler, common as a final global step) and the modern alternative of strided convolutions doing the downsampling themselves. The tool varies; the principle, spend resolution to buy abstraction, is permanent.
Check your understanding
What does max pooling actually discard, and what does it keep?
What to remember
- ✓After detection, precise position is nearly worthless, presence and strength of evidence are what count.
- ✓Max pooling: keep each 2×2 window’s champion. Map halves per side; strong detections survive; small shifts barely register.
- ✓Three dividends: cheaper compute, wider receptive fields for later layers, built-in shift tolerance.
- ✓Across the network: channels grow, resolution shrinks, WHERE is continuously traded for WHAT.
- ✓Conv, ReLU, pool, you now hold all the bricks. Next lesson: the architectures that stacked them into history.