19 · Images: photographs

Thursday, Oct 29, 2026

Materials for this session are not published yet. They appear here before class.

Objectives

By the end of this session you can:

Frame a photos-to-variable pipeline.
Compute a feature from raw pixels yourself — a photo is a grid of numbers.
Score images with a vision model, or with embeddings plus a simple classifier.
Validate scores against ground truth on photos the model never saw — and say which method won, by how much.
Articulate the ethics constraints that face data demands.

What photographs have measured: pollution from haze, crowd sizes, emotion in public life.
The full arc on ordinary photos: labels → model → score → validate.
First, the baseline you build yourself: a photo is a grid of numbers — average the sky’s pixels for color, take the variance for contrast, and you have a haze score in thirty lines of plain R.
Then the model routes: zero-shot scoring with a vision model; embeddings plus a simple classifier; fine-tuning as a concept.
The contest, measured: the vision model against your thirty lines, both scored on held-out photos against the matched monitor readings. The model should win — the interesting number is by how much.
Live demo: emotion classification on faces, scored against a gold set labeled in class — with the ethics conversation face data demands.
The flagship dataset: 365 daily pollution photos, one per day for a year, matched to same-day PM2.5 readings from a monitor.

Verification habit. Hold out labeled photos, report agreement, and inspect the failures.