Explain like I'm five · AI Theory

What is a CNN (Convolutional Neural Network)?

CNNs, told as a story about an eye that scans a small window bit by bit, instead of staring at the whole photo at once.

trying to take in the whole photo at once — there are just too many pixel numbers

Trying to take in the whole photo at once — there are too many pixel numbers.If the sorter reads every pixel separately, the computation gets out of hand.

where the seed-mark pattern sits — invisible when you only see loose numberstaking in everything as loose numbers makes local patterns slip right past

With pixels laid out as loose numbers, you lose where the local patterns sit.You have to keep track of each pixel's position to ever see the pattern.

a CNN slides a small window bit by bit, finds local patterns, and stacks them to judge the whole pictureoriginal photosmall pattern map (feature map)the filter nudges the window one cell at a timeonly where a pattern was found stays bright

A CNN slides a small window bit by bit, finding local patterns.A CNN stacks the patterns it finds across layers and judges what the whole photo shows.

Slide a small windownudging one cell at a time, hunting for a pattern
Keep only what mattersoften keeping the strongest signal (pooling)
Stack several layerssmall patterns build into bigger ones
Classify at the endmaking the final call from everything stacked up
FOUR STAGES OF A CNNconvolution — a small window finds local patternsactivation — adds nonlinearity so complex patterns can be learnedpooling — shrinks the map, often keeping the strongest signalfully connected — makes the final classification from it all

Convolution, activation, and pooling repeat in layers; a fully connected layer typically classifies at the end.Layer by layer, tiny patterns like dots and lines grow into bigger ones, like the shape of an apple.

CNNs are strong with photos and patternsa CNN doesn't 'understand' the way a person doesbecause a CNN relies on pixel patterns, a few nudged dots can fool it

A CNN is strong with photos and patterns — but it doesn't understand them like a person.Which is why a CNN's call can still be flipped by tiny tampering like adversarial attacks.

In one breath

A CNN = a neural network that scans a photo through a small window, bit by bit, finding local patterns and stacking them to judge the whole photo.

A CNN (Convolutional Neural Network) is built around the convolution operation: sliding a small filter (kernel) across the whole input to find local patterns — edges, color patches, and the like. Stacking convolution, activation, and pooling across many layers lets shallow layers recognize simple patterns while deeper layers recognize more complex shapes. CNNs are especially strong at image and video classification, but because they rely on pixel-level patterns, they remain vulnerable to subtle perturbations like adversarial examples.

When grown-ups say it

CNN
The eye that scans through a small window, bit by bit. Built around the convolution operation.
Convolution
The operation of sliding a small filter bit by bit. The core move that hunts down local patterns.
Filter (kernel)
The small dashed window that hunts for a pattern. Many different filters exist, each hunting a different kind of pattern.
Feature map
The small map where only pattern-found spots stay bright. What comes out after one pass of convolution.
Pooling
Shrinking the map, often keeping the strongest signal. Cuts computation and makes the call less rattled by small shifts in position.
Layer
The stage where small patterns build into bigger ones. Shallow layers see simple patterns; deep layers see complex shapes.
Fully connected layer
The layer that gathers everything stacked up and makes the final call. Typically sits at the very end of a CNN, deciding the bin.
Adversarial example
An attack that exploits the weakness of reading pixel patterns alone. A CNN, leaning on pixel patterns, can still be fooled by this. → The Fake Dot Pattern Snuck Onto the Produce
Next story
Feeding the Scale's Number Through an S-Curve →
Want another term as a picture book? Request →