Convolutional Neural Network
A convolutional neural network slides small filters across its input to detect local patterns, which is what made it the standard approach for images.
Machine Learning
Key takeaways
- A CNN slides small learned filters across its input to find local patterns.
- Weight sharing means a pattern learned in one place is recognised everywhere, with far fewer parameters.
- Layers stack into a hierarchy: edges, then textures, then objects.
- It works on anything with local structure, not just images.
A convolutional neural network slides small filters across its input, looking for the same local pattern everywhere. It is the architecture that made image recognition work.
The trick is weight sharing. One filter that detects an edge is reused across the whole image, so the network learns far fewer parameters. A pattern learned in one corner is recognised in any other.
Think of it like this. Think of moving a small stencil across a page, checking everywhere for the same shape. You do not need a separate detector for every position, because one stencil works anywhere you put it.
How it works
- Filters slide across the input. Each one is a small grid of learned weights, typically 3x3.
- The same filter is used everywhere. This is weight sharing, and it is why a CNN needs far fewer parameters than a fully connected network.
- Layers build hierarchy. Early filters find edges, later ones find textures, then parts, then objects.
- Pooling shrinks the picture between layers, keeping the strongest signals and reducing computation.
Common mistakes
- "CNNs are only for images." Anything with local structure works: audio spectrograms, time series, even text.
- "Bigger filters see more." Stacking several 3x3 layers covers the same area with fewer parameters and more nonlinearity, which is why almost everything uses small filters.
- "Pooling is mandatory." Many modern architectures use strided convolutions instead, or skip it entirely.