Convolutional Neural Networks (CNNs) Cheat Sheet

Convolutional Neural Networks are a class of deep learning models designed to process data with a grid-like topology — most famously images — by exploiting local spatial correlations through learnable filters. They sit at the heart of modern computer vision, enabling tasks from image classification and object detection to medical imaging and autonomous driving. Unlike fully connected networks, CNNs achieve parameter sharing and translation equivariance by sliding the same filter across the entire input, which dramatically reduces parameter count while preserving spatial structure. The key mental model to hold throughout is that a CNN is a hierarchy of feature detectors: early layers learn edges and textures, middle layers learn parts, and deep layers learn semantics — and every design choice (kernel size, stride, normalization, skip connections) shapes how information flows through that hierarchy.