CNN Building Blocks, Receptive Fields, and Residual Design Patterns

A chart-image classifier and a volatility-surface heat-map model share the same design question: how much spatial context should each layer see, and how deep can the stack grow before gradients or resolution collapse the training run? After convolutional foundations, stacking more layers is not automatically better — each block must widen its receptive field, preserve useful resolution, and remain trainable. This note develops receptive-field arithmetic, standard convolutional building blocks, and residual skip connections as reusable design patterns rather than named-architecture trivia.