What it is
A convolutional neural network is a model built for grid data such as images. Instead of connecting every pixel to every neuron, it slides small filters across the input and reacts to local patterns like edges and corners.
The convolution operation
A filter is a small grid of weights, often three by three. The network slides it over the image and computes a dot product at each position, producing a feature map.
- A filter learns to fire on one kind of pattern, such as a vertical edge
- The same filter weights are reused at every position, which is called weight sharing
- Many filters run in parallel, so one layer can detect many pattern types
Why it works
Weight sharing gives two big wins.
- It uses far fewer parameters than a fully connected layer on the same image
- It gives translation invariance, so a pattern is recognized wherever it appears
Stacked convolution layers build a hierarchy. Early layers find edges, middle layers find textures and shapes, and deep layers find whole objects.
Key idea
A convolutional network reuses small filters across an image to detect local patterns with few parameters.