You already know: convolutions, CNNs and the idea of learned features (D4 Major architectures); training loops and transfer of representations (D4, D6); attention and transformers (D5); precision, recall and thresholds (D3 ML evaluation); the D17 bridge's cost of an image in tokens.
Why this matters for Atlas. Vision is the most mature non-text modality and the front end for documents, diagrams, screenshots, video and robotics. This lesson builds it from the ground up: images as arrays and convolution as a local pattern detector; why CNNs work and why vision transformers (ViTs), which treat an image as a sequence of patches, now dominate; transfer learning, the way almost every real vision system is built; and the two workhorse tasks beyond classification, object detection (boxes, IoU, non-maximum suppression, mAP) and segmentation (pixel masks). Vision-language models get their own lesson next.
1. Images are arrays; convolution finds local patterns
The question. How does a computer see an edge in a photo?
Start from nothing. A colour image is an array of shape height × width × 3 (red, green, blue), each value 0–255. A convolution slides a small kernel (say 3×3) over the image, computing a weighted sum at each position: a local pattern detector. A kernel with negative weights on the left and positive on the right responds strongly where brightness jumps from left to right: a vertical edge. Classical vision hand-designed such filters (Sobel edges, HOG gradient histograms); CNNs learn thousands of them, stacked so that early layers detect edges, middle layers textures and parts, and late layers objects.