A digital image is a grid of numbers, and a convolution is a small window of weights slid across that grid, multiplying and summing at every position. From this one simple operation, deep networks build their entire visual vocabulary, starting with edges and ending with objects.
Filters as Pattern Detectors
Each filter is a tiny template. A filter with positive weights on the left and negative on the right responds strongly wherever a vertical edge appears, producing a feature map that lights up at those locations. Classical vision engineers designed such filters by hand for decades; a convolutional network instead learns thousands of them from data, discovering whichever patterns best serve the task.
Why Sharing Weights Matters
The same filter is applied at every position in the image, which has two profound consequences. First, the network needs dramatically fewer parameters than a fully connected layer over raw pixels, making training tractable. Second, it gains translation equivariance: a cat detector that works in the top-left corner works identically in the bottom-right, because it is literally the same weights looking.
Depth Builds a Hierarchy
Stacking convolutional layers composes simple detectors into complex ones. Early layers respond to edges and color blobs, middle layers combine those into textures, corners, and simple parts, and deeper layers assemble parts into faces, wheels, or wings. Visualizing what activates each layer is one of the best exercises for building genuine intuition about what your network has learned.
Action Step
Using any deep learning framework, apply a hand-written vertical-edge filter and a horizontal-edge filter to a photograph and display the two resulting feature maps. Confirm each map highlights exactly the edges you predicted it would.
Educational content only. The AI/ML tooling landscape changes quickly — verify current library versions, APIs, and best practices in official documentation. This course does not certify anyone.