A 224×224 RGB image is 150,528 numbers. A fully connected layer on that input would need tens of millions of parameters for the first layer alone. Convolutional neural networks fix this with local filters, weight sharing, and a hierarchy of features. This chapter moves from the basics—convolutions and receptive fields—through ResNet to modern tasks: detection, segmentation, and diffusion-based generation. PyTorch + torchvision is the practical stack.