Vision Introduction
Convolution
Reuse a small filter across a larger input.
Before this
The intuition
A filter moves across an input, combining a small neighborhood at each position. Sharing the same weights lets it detect a pattern in different locations.
A concrete example
For input [1, 2, 4] and a two-entry filter [1, −1], sliding without flipping gives [−1, −2]. This cross-correlation convention is commonly called convolution in neural networks.
What to remember
Filter size, padding, and stride determine the output shape. A filter detects local patterns; it does not understand an image by itself.
Check your understanding
How many positions fit a width-2 filter into a width-4 input with stride 1 and no padding?
Reveal explanation
3 positions.