Unit at a Glance: Picture This: Image Classification + Data Bias

PICTURE THIS · TEACHER GUIDE

Picture This: Image Classification + Data Bias

A complete five-day, teacher-led sequence with whole-group decks and ready-to-print student materials.

Topic: Model training, testing, and revision

Student objective: Students will Train a classifier, test unseen examples, identify a limitation, connect a prediction to code, and revise

Essential TEKS:

§126.19(c)(3)(B) — discuss and implement a design process that includes planning, selecting digital tools to develop, test, and evaluate design limitations, and refining a prototype or model; and

Supporting TEKS:

§126.19(c)(2)(C) — modify and implement previously written code to develop improved programs.
§126.19(c)(6) — The student uses digital tools to transform data, make inferences, and predictions. The student is expected to use digital tools in order to transform data, analyze trends, and predict possibilities and develop steps for the creation of an innovative process or product.

Demonstration of learning: model tests, planning work, and remix/share.

Alignment status: D

Topic: Image classification, training data, model failure, and responsible AI

Objective: Students will move from sorting images by hand to training an image classifier and coding with its predictions.

Demonstration of learning: Students submit a working .sb3 project and explain one model failure and one revision that improved it.

Day	Learning move	Student evidence
1 · Meet the AI	Train thumbs-up/thumbs-down, then deliberately try to break the classifier.	Two trained classes + a named failure condition.
2 · Data makes the model	Independent packet: collect, train, test; then connect missing variety to data bias.	Annotated three-page packet and five responses.
3 · Be the AI	Sort cat/dog images by hand, write rules, recreate the classifier, then test it with images that expose its gaps.	Rules, test results, and exit explanation.
4 · Train + connect	Design two or three classes, train with variety, partner-test, export, and connect to RAISE Playground.	Planning sheet, model link, and working response.
5 · Remix + critique	Change the sprite/actions/classes, retest for failures, revise, share, and submit.	.sb3 upload + short reflection.

The slide decks are the whole-group teaching route. Canvas pages are the student directions and absence route.

Day 1: Download Day 1 deck · Make a Google copy

Day 2: Packet PDF Editable packet · Make a Google copy

Day 3: Download Day 3 deck · Make a Google copy Cat/dog datasets Bias test photos Exit ticket

Day 4: Download Day 4 deck · Make a Google copy Planning sheet PDF Editable planning sheet · Make a Google copy

Day 5: Download Day 5 deck · Make a Google copy Exit ticket

Test webcam permission on a student Chromebook in Teachable Machine.
Open the RAISE Playground once on a student device; the first load can be slow.
Print Day 2 one packet per student.
For Day 3, print pages 1 through 7 of the cat/dog dataset packet per pair, page 1 of the bias-test packet per table, and one exit ticket per student. Add scissors and glue or tape. Keep the final credits pages with the digital files.
For Day 4, print one planning sheet per pair. Students upload a photo, scan, or digital copy of the sheet and paste the model URL in the Canvas submission comment so they can reopen it on Day 5.
The final turn-in is an .sb3 file. Reserve the last five minutes for File → Save to your computer, then Canvas upload.

Day 1: teach slides 1-9, then leave the build/help screen up while students train and test.
Day 2: project the Canvas page or packet directions; this day is intentionally independent/sub-friendly.
Day 3: move quickly through the opening sort so the final 14 minutes remain for the Teachable Machine bias test.
Day 4: leave the training checklist up during work time; students should get one response working in Playground before the bell.
Day 5: leave the remix menu and success criteria up; stop with enough time for a quick share and file submission.

A strong explanation names the missing pattern in the training data, not just 'the AI was wrong.'
Class names in Teachable Machine and Playground must match exactly.
More photos are not automatically better. Students need useful variety: people, distance, angle, lighting, and background.
The model makes a prediction; the student-written code decides what happens next. Keep those two ideas separate.

RAISE Playground was created by the MIT RAISE Initiative and the Personal Robots Group at the MIT Media Lab.