Seeing Images

A CNN processes an image with small filters, typically 3x3 grids of weights. Each filter slides across the whole image, and at every position it multiplies the pixels underneath by its weights and adds them up — the operation you see here, called convolution. A single filter detects one feature, like a vertical edge. A CNN uses many filters per layer, and stacks layers: the first layer finds edges, the next combines edges into corners and textures, later layers combine those into shapes and eventually whole objects like a cat's ear or a wheel. Crucially the same filter is reused across the entire image, so the network needs far fewer weights than a fully connected one, and it recognises a feature no matter where it appears. This weight-sharing is what made CNNs the workhorse of image recognition.

CNN history is a lineage of depth. LeNet (1998) read bank-cheque digits with two conv layers. AlexNet (2012) went deeper on GPUs and won ImageNet by a landslide — the moment deep learning took over vision. VGG showed that stacking many uniform 3x3 filters beats fewer big ones. ResNet (2015) added skip connections that let gradients bypass layers, unlocking networks 100+ layers deep — the residual idea now lives inside transformers too. EfficientNet scaled depth, width, and resolution in a balanced ratio for the best accuracy per FLOP. Since then Vision Transformers have taken much of the crown, but CNNs still rule where compute is tight: phones, cameras, embedded devices.