Skip to content
Learn Motion
ExploreHow it worksMembership
Log in
Learn Motion

How AI Sees: A Conceptual Guide to Image and Video Understanding

1What a Machine Actually Sees2From Pixels to Patterns: Features3How Deep Networks Learn to See4Beyond Labels: Locating and Describing What Is Seen5Adding Time: Understanding Video6How These Systems Learn and How We Judge Them
How Deep Networks Learn to See

A Filter That Slides

2 / 5
Watch the small grid move across the image. At every position it multiplies its weights by the pixels underneath and writes a single number into the map on the right. Where the patch matches what the filter is tuned for, that number is large; where it does not, the number is small. The crucial point is that the weights are learned, not written by a person. They start random and get adjusted until the filter becomes useful, so a filter may end up looking like an edge detector, or a color-transition detector, or something no one would have thought to design. And because the same nine weights are reused at every position, the detector fires wherever its pattern appears, top left or bottom right. That is position tolerance built into the operation itself.
0:00 / 0:00

A convolution filter is a small grid of weights — three by three, five by five — that is dragged across the image. At each position it multiplies its weights by the pixels underneath, sums the products, and writes one number into an output map. That output map is called a feature map, and a large value in it means the patch under the filter looked like whatever the filter is tuned to detect.

The weights are not designed by a person. They start random and are adjusted during training until the filter becomes useful, which is the decisive difference from the hand-designed edge kernels of the previous chapter. A filter that ends up with positive weights on one side and negative weights on the other behaves like an edge detector; a filter with a different arrangement may respond to a color transition or a small blob.

Because the same weights are reused at every position, two consequences follow. First, the number of parameters stays tiny — nine weights cover the whole image rather than one weight per pixel. Second, the detector is position-tolerant by construction: if the pattern appears in the upper left or the lower right, the same filter fires. That is exactly the property a hand-designed recipe could not guarantee.

Previous2 / 5Next

Learn Motion

Generate a course. Learn it properly.

Operated by Wuhan Daoyin Technology Co., Ltd.

Contact: [email protected]
Privacy PolicyTerms of Service

© 2026 Learn Motion