AI Technical AptitudeAI Methods9-12.AITA.AIM.5

Three Families, Three Jobs

9–12 CS / engineering50 minutes90 minutes (lab extension)

Standard quoted exactly

Understand the differences between common machine learning models.

Example from the standards. Common models include logistic regression, decision trees, and neural networks.

Student-friendly learning targets

I can describe logistic regression, decision trees, and neural networks in high-school language.
I can match each family to a job it is good at and a job it is a poor fit for.
I can compare the three on interpretability, data hunger, and compute without training any of them.

Essential questions

If they all 'learn from data,' why not use one family for everything?
What does a tree know how to show that a neural net usually does not?
Why is logistic regression still in the toolkit?

Objectives

Build a three-column comparison: how it decides, what it outputs, what it needs, what it cannot do (preview of AIM.6).
Walk a 20-questions tree on paper, a logistic probability on a frost yes/no, and a neural net as stacked simple units for pixels.
Give each family an Idaho job.
Avoid treating 'neural net' as a synonym for all of AI.

Key vocabulary

Logistic regression: Draws a linear boundary and turns the distance to that line into a probability of yes vs. no. Interpretable coefficients; fails XOR (DA.6 / AIM.6).
Decision tree: A flowchart of questions (Is RH < 15%? Is wind > 20 mph?). Easy to read; can overfit; can show which features get used.
Neural network: Layers of weighted sums plus simple nonlinear squashes. Flexible enough for images and language; data- and compute-hungry; hard to explain.
Interpretability: Whether a person can see why the model said yes. Trees high, logistic medium, deep nets low.
Overfit: Memorizing the training rows so the model looks brilliant at home and fails on new nights, counties, or photos.

Teacher background

The standard asks for differences among common models; the example names the three families. Stay conceptual. Logistic: one line + probability — good for frost yes/no on a logger. Tree: 20 questions for crop diagnosis — good when experts already think in branches. Net: image of a plume — good when the pattern is texture, not three numbers. Differences to hit: decision surface (line vs. axis-aligned splits vs. flexible), interpretability, data/compute, typical jobs. Do not derive backprop. Paper tree, paper line, stacked-boxes net. Spreadsheet can compute a logistic sigmoid of a linear score if you want a number, not required. AIM.6 will stress limitations (XOR, feature importance); today is family identity.

Materials and prep

Materials

Three family posters with a blank 'how it decides' box.
Paper 20-questions tree for a 6-leaf crop-diagnosis toy (public symptoms, no student health).
Logistic handout: a linear score (a*temp + b*humidity + c) then a 0–1 squish described in words.
Neural-net stacked boxes: pixels → hidden → label, with a 'many knobs' caption.
Comparison table (print).

Before class

Draw a tree that uses RH, wind, and leaf spots so it feels Idaho.
Keep XOR off-stage except a one-line 'logistic is linear; details tomorrow in AIM.6.'

Instructional sequence

Name the tool on the bench

5 min

Three unlabeled pictures: a line through points, a flowchart, a stack of layers. Students guess names.
Reveal logistic, tree, neural net. Announce: differences, not training.

How each one decides

12 min

Logistic: weighted sum of features, squashed to a probability, cutoff at 0.5 (or another threshold). Job: frost yes/no on few features.
Tree: ask a feature question, follow yes/no, reach a leaf label. Job: 20 questions for 'which deficiency.'
Net: many such weighted sums in layers; good when inputs are pixels or tokens. Job: smoke vs. cloud on a photo.
Fill interpretability, data hunger, compute on the board table.

Walk each family once

10 min

Run one new 'night' through the logistic score on paper (teacher supplies weights).
Walk one diagnosis through the tree by answering questions aloud.
For the net, trace a cartoon photo through two layers as 'numbers in, numbers out' — no algebra slog.

Comparison table

10 min

Students complete the three-column table and assign one Idaho job and one poor-fit job per family.
Poor-fit examples: logistic on raw photos; a giant net on 12 rows of GDD; a 10,000-node tree a grower cannot print.

Real-world examples

Logistic: binary 'will this canal exceed a legal cutoff tonight' on a low-power station.
Tree: extension 20-questions for potato nutrient issues used as a phone script.
Neural net: classifying GOES fire detections or counting cars in a public ITD camera (not student parking lots).
Sports: logistic win-probability from score and time; a tree of 'foul trouble' questions; a net on video of a shot — three families, one game.

Hands-on activity

Family speed round

8 min

Teacher reads five jobs; teams flash a card: L, T, or N, plus a 5-word why.
Include one trick (XOR or a GDD formula) that should be 'none — use a rule/procedure' to keep AIM.4 alive.
Correct as a class without shaming; the point is distinctions.

Discussion questions

Why would a county commissioner prefer a tree they can argue with?
When is logistic's simplicity a safety feature?
What does a neural net's extra power cost you in data and explanation?
Are LLMs a kind of neural net? (Yes, huge ones — still not the right tool for XOR.)

Differentiation

Support

Table mostly filled; students add jobs only.
Physical walk along floor tape as a tree.

Challenge

Add 'random forest as many trees' in one paragraph without claiming it is a fourth required family.
Explain sigmoid in words: large positive score → probability near 1.

Multilingual learners

Family names + icon; jobs written with sentence stems.
Allow the table's 'how it decides' in the student's strongest language.

IEP / 504

One family mastered deeply (tree) plus headlines for the other two if processing is slow.
No required sigmoid math.

Assessment

Formative

Warm-up picture naming.
Correct walk of the paper tree.

Summative

Completed comparison table with a job and a poor-fit per family.
Speed-round participation as supporting evidence.

Success criteria

Student describes how each family decides in ordinary language.
Student does not call every model a neural net.
Student matches at least two families to fitting Idaho jobs.

Responsible use, ethics, and privacy

Responsible use

Understanding families is so you stop buying 'an AI' and start asking which family, which job, which data.

Ethics

An uninterpretable net on a high-stakes yes/no (loan, discipline, medical) may be the wrong family even if it scores well.

Privacy

Toy trees use public crop symptoms. No student health, attendance, or discipline rows to 'compare models.'

Reflection

Which family would you trust to explain itself to a grower?
Which family would you pick if you had 40 rows vs. 40,000 photos?
What difference will you remember in six weeks?

Homework

Complete the take-home comparison: in your own words, how each family decides; one Idaho job; one poor fit. Five sentences total plus the table. No training.

Closing

Three families, three jobs: a probability line, a question flowchart, a layered pixel-eater. They are not interchangeable. Tomorrow we ask what each one cannot do — including XOR for the linear family.

Extensions and cross-curricular links

Go further

90-minute block: build a larger paper tree from a 12-row crop table (still no software); compare its questions to logistic's weights list.
Python extension: train all three on a tiny public table and print accuracy — only after AIM.9 if possible, or with a 'accuracy may lie' warning.
Card-sort properties (interpretable, data-hungry, linear, pixel-friendly) onto families.
AIM.6 next will ask what each cannot do; leave XOR as a cliffhanger.

Biology: Dichotomous keys are decision trees; students already know the shape.
Algebra: Logistic starts from a linear combination students recognize.
Media arts: Image models are nets because photos are grids of numbers, not three features.