Convolutional Networks for Structured Data

A fully connected MLP treats every input feature as unrelated to its neighbours. That assumption is expensive and often wrong for image-like inputs, where nearby pixels form edges, textures, and objects. Convolutional neural networks (CNNs) encode the right structural prior: a small learned filter is reused across every location in a grid. This note develops local receptive fields, parameter sharing, stride, padding, and pooling, then shows why these choices make CNNs a practical architecture for vision and for grid-structured quantitative data.