{
  "$type": "site.standard.document",
  "bskyPostRef": {
    "cid": "bafyreihuezhnkrwm4y2v7hdkluanom5udnskun5fzeitzv27oirrtvf54m",
    "uri": "at://did:plc:25rdn5elo5izoxrmtis34zuk/app.bsky.feed.post/3mope6nzn54g2"
  },
  "coverImage": {
    "$type": "blob",
    "ref": {
      "$link": "bafkreiclhjhusq2rqn46mbqnhge6etddwdrpwu47eb67ocsxov566lfply"
    },
    "mimeType": "image/webp",
    "size": 64354
  },
  "path": "/dev48v/pooling-in-cnns-shrink-the-map-keep-what-matters-1p74",
  "publishedAt": "2026-06-20T07:02:56.000Z",
  "site": "https://dev.to",
  "tags": [
    "machinelearning",
    "deeplearning",
    "ai",
    "beginners",
    "https://dev48v.infy.uk/dl/day9-pooling.html",
    "Pool a grid."
  ],
  "textContent": "After a few conv layers you're drowning in feature maps — too big and slow. Pooling shrinks them while keeping the signal, and it has zero parameters. Four numbers in, one out.\n\n🪟 **Max vs average pooling:** https://dev48v.infy.uk/dl/day9-pooling.html\n\n##  The operation\n\n\n    for (let y = 0; y < h; y += 2)        // stride 2: hop, don't slide\n      for (let x = 0; x < w; x += 2) {\n        const win = [img[y][x], img[y][x+1], img[y+1][x], img[y+1][x+1]];\n        out[y/2][x/2] = reduce(win);       // 2×2 → 1 value, halves both dims\n      }\n\n\n##  Max pool — keep the strongest\n\n\n    const reduce = (w) => Math.max(...w);\n\n\nA feature detector asks \"is this feature here?\" — max pooling answers \"yes, somewhere in this region it fired\", keeping the feature's presence while discarding its exact location.\n\n##  Average pool — smooth it\n\n\n    const reduce = (w) => w.reduce((a,b)=>a+b) / w.length;\n\n\nCommon at the end of a network (global average pooling) to summarise each map before the classifier.\n\n##  The real point: translation tolerance\n\nBecause pooling reports \"feature present in region\", shifting the input a pixel barely changes the output. That invariance is why a CNN recognises a cat whether it's top-left or centre.\n\nAnd it's free — no learnable weights. (Modern nets sometimes use strided convolutions instead, but the intent is identical.)\n\n##  The takeaway\n\nHop in 2×2s, keep the max → smaller maps, position-tolerant, zero params. Pool a grid.",
  "title": "Pooling in CNNs: Shrink the Map, Keep What Matters"
}