THE VR SCHOOL
  • Schools & Districts
    • License the academy
    • Free program for your students
    • Request a demo
  • Catalog
    • AI & Data
    • Cybersecurity & Web
    • Programming
    • Health Science
    • All Guilds
    • All courses
  • Families
  • Accreditation
  • Log In
  • Request a demo
  • Home
  • Courses

RLP AI Design A — Machine Learning, Media and the Human Body

Curriculum

  • 4 Sections
  • 20 Lessons
  • Lifetime
Expand all sectionsCollapse all sections
  • Unit 1: How Machines Learn: Making AI Visible
    5
    • 1.1
      Explaining the Invisible: Designers Who Made Data Visible
      50 mins
    • 1.2
      Sorting by Features: Training a Classifier by Hand
      100 mins
    • 1.3
      Critiquing AI Explainers: Accuracy, Clarity and Hype
      50 mins
    • 1.4
      Designing the Explainer: Hierarchy, Sequence and Storyboard
      100 mins
    • 1.5
      Portfolio Task — How a Machine Learns: Explainer and Storyboard
      200 mins
  • Unit 2: Human-Centered AI: History, Ethics and Authorship
    5
    • 2.1
      Drawing Time: Timeline Design and the History of AI
      50 mins
    • 2.2
      Verified Event Cards: Building a Physical Timeline Storyboard
      100 mins
    • 2.3
      Who Is Responsible? Bias and Human-Centered AI in Medicine
      100 mins
    • 2.4
      Authorship and AI Imagery: Making the Presentation Yours
      150 mins
    • 2.5
      Portfolio Task — The Evolution of a Healthcare AI Innovation
      200 mins
  • Unit 3: Drawing the Body: The Eleven Organ Systems
    5
    • 3.1
      Art and Anatomy: Five Centuries of Seeing Inside the Body
      50 mins
    • 3.2
      Sculpting a System: Modeling Structure and Function
      100 mins
    • 3.3
      Clinician and Patient: Role-Play and the Critique of Health Communication
      100 mins
    • 3.4
      Illustrating a Condition: Panels for One Body System
      150 mins
    • 3.5
      Portfolio Task — The Body Atlas Exhibition
      200 mins
  • Unit 4: Signals and Diagnosis: Visualizing Vital Signs and Medical Data
    5
    • 4.1
      Data Portraits: Du Bois, Einthoven and the Art of the Signal
      50 mins
    • 4.2
      Pulse to Pixels: Coding a Heart-Rate Visualization
      100 mins
    • 4.3
      Who Sees Your Data? Medical Records, Genetics and HIPAA
      50 mins
    • 4.4
      Imaging the Invisible: Designing a Patient Brochure
      150 mins
    • 4.5
      Portfolio Task — Reading the Body’s Signals
      200 mins

Sorting by Features: Training a Classifier by Hand

Unit 1  ·  Studio Techniques Lab  ·  Lesson 2 of 20

Sorting by Features: Training a Classifier by Hand

MA:Cr1A.IRLP-AIDA
Hands sort round and crescent cell cards while comparing a hand rule with an abstract machine model.
By the end of this lesson I can…

explain supervised learning, features, labels and the training-testing split by building and testing a classification rule by hand, and compare my rule with a machine-trained model.

Instruction

Most AI used in medicine today is built with machine learning: instead of a programmer writing every rule, the system adjusts itself to fit examples. The most common form is supervised learning. You give the system many examples that have already been labeled by people (this retinal photograph shows disease, this one does not), and it learns a way to map the input to the label. The labeled examples are the training data.

Features. A model does not see a picture the way you do. It works from measurable properties called features. Older systems used features people chose: the length and width of a cell, its brightness, the number of edges. Modern deep-learning systems learn their own features from raw pixels, which is powerful but makes them harder to explain. Either way, the model can only learn from what is in the data.

A worked example. OpenStax Anatomy and Physiology 2e (section 18.3) describes the normal red blood cell as a biconcave disc: round, with a thinner, paler center. In sickle cell disease an abnormal hemoglobin causes many red cells to take a crescent or sickle shape. Suppose we want to sort drawings of cells into “round” and “sickle-shaped”. One feature is the length-to-width ratio. A round cell 8 mm long and 8 mm wide in a drawing has a ratio of 1.0. A sickle cell 16 mm long and 4 mm wide has a ratio of 4.0. A simple rule might be: if the ratio is greater than 2.0, label it sickle-shaped. That rule is a tiny model. The number 2.0 is its threshold, and choosing it is a form of training.

Classification and regression. Our rule outputs a category, so it is a classifier. A model that outputs a number (predicting a patient’s heart rate, or how many days a hospital stay will last) performs regression. You will use both words all year.

Testing on unseen data. If you set your threshold by looking at twenty cards, the rule will fit those twenty well. The real question is how it performs on cards it has never seen. That is why developers hold back a test set. A rule that fits its training data perfectly but fails on new data is overfitted; it has memorized quirks instead of learning the pattern.

Why a hand model first? When you later build a second model on the same cards, by hand or (with a parent’s permission) in a no-account browser tool such as Teachable Machine, you will understand what any trained model is doing: finding features and boundaries that separate the labeled examples. You will also understand its weaknesses. If every sickle card in the training set was drawn in red marker, the machine may learn “red” rather than “crescent”. Your drawings are data, and data carries the habits of whoever made it.

Studio habit. Keep a process sketchbook page: your features, your threshold, your errors and a small diagram of the boundary. Media artists document process; it becomes material for your explainer.

A measured card set, if you cannot draw and measure forty cards. The forty drawings below were measured for this course; the numbers are illustrative, not taken from any patient. Thirty are the TRAIN set, with their labels shown, and ten are the TEST set, with the labels held back to the end of this section so you can predict before you check. Compute the length-to-width ratio yourself, to one decimal place, exactly as you would with your own cards.

TRAIN (30 cards, labels shown).

CardLength (mm)Width (mm)True label
C01246sickle
C0286round
C03118round
C04123sickle
C05113sickle
C06116sickle
C07104sickle
C0887round
C091010round
C10195sickle
C11134sickle
C1289round
C13166sickle
C14164sickle
C1587round
C1686round
C17145sickle
C18196sickle
C191210round
C2077round
C21156sickle
C221010round
C23109round
C2483sickle
C2577round
C26185sickle
C27155sickle
C2867round
C291310round
C3076round

TEST (10 cards, labels held back).

CardLength (mm)Width (mm)
C31133
C32119
C33144
C34216
C351412
C36175
C3776
C38178
C3986
C40185

TEST labels — do not read these until your ten predictions are written down. C31 sickle; C32 round; C33 sickle; C34 sickle; C35 round; C36 sickle; C37 round; C38 round; C39 round; C40 sickle.

Two of the forty are deliberately awkward, and that is the point: no single threshold gets all forty right, and finding where your rule breaks is the lesson.

📖 Read for this lesson

The reading opens over the lesson; close it to come back to exactly where you were. The checks below draw on it.

  • Forty Measured Cell Cards: The Data for the Hand-Built Classifier LabSLS original teaching data set and worked solution; no external source. Teachable Machine (Google) is named as an optional tool only; retrieved Sept. 26, 2026 · about 7 min

    The measured card set the lab’s no-computer route calls for, with TRAIN and TEST tables, the steps for the two models, the color-shortcut trap, and an answer key you open only after you have made your predictions.

Vocabulary in context

  • machine learning — Building a system that improves at a task by adjusting itself to fit examples rather than following only hand-written rules.
  • supervised learning — Machine learning from examples that people have already labeled with the correct answer.
  • feature — A measurable property of an input, such as length-to-width ratio, that a model uses to decide.
  • test set — Labeled examples held back from training and used only to measure how well the model works on unseen data.
  • overfitting — Fitting the training examples so closely that performance on new data gets worse.

Studio: Paper cells and a hand-built classifier

Question. Can a single measurable feature separate drawings of round and sickle-shaped red blood cells, and how does a hand-written rule compare with a machine-trained model on cards neither has seen?

Materials

  • Forty cards, which you draw and measure yourself. Choose one: (1) blank index cards; (2) rectangles cut from cereal-box card or any packaging; (3) forty squares of plain paper, which work exactly as well because only the drawing is measured.
  • Two colors to draw with, because the color shortcut is part of the experiment. Choose one: (1) fine-tip water-based markers in black and one color; (2) a blue pen and a red pen; (3) a pencil and any one colored pencil or crayon.
  • A millimeter scale. Choose one: (1) a clear plastic ruler; (2) any ruler with a metric edge; (3) graph paper with 5 mm squares, counting squares and halving.
  • Two containers labeled TRAIN and TEST. Choose one: (1) two envelopes; (2) two folded sheets of paper; (3) two marked areas of a table.
  • Sketchbook or plain paper for the process page.
  • For the second model, choose one: (1) no computer at all, using the by-hand two-feature model described in the lab’s alternative, which is the default route; (2) a free spreadsheet, plotting the TRAIN cards as a scatter chart and drawing the same boundary line on it; (3) only with a parent’s or guardian’s permission, Teachable Machine on any computer, which trains in the browser and asks for no account (use a webcam, or upload phone photographs of the cards as image files).
Safety.
  • Use water-based markers only; permanent markers need ventilation and are not needed here.
  • Do not share personal medical information; the cells are drawings, not anyone’s diagnosis.
  • If you use a computer, keep its cables out of walkways.

Procedure

  1. Draw 20 round cells (with a pale center) and 20 sickle-shaped cells, varying size, color and pen pressure on purpose. Label each card on the back only. Draw all 40 yourself.
  2. Shuffle the cards. Put 30 in the TRAIN envelope and 10 in the TEST envelope without looking at the labels.
  3. Measure the length and width of every TRAIN card, calculate the length-to-width ratio to one decimal place, and record it with the label.
  4. Choose a threshold that best separates the two labels in the TRAIN data. Write the rule as a sentence: If ratio > __, label sickle.
  5. Apply the rule to the 10 TEST cards before turning them over. Record each prediction, then check the labels and count errors.
  6. Build the second model. By default, use the by-hand two-feature model in the alternative (or the same boundary line drawn on a spreadsheet scatter chart) and record its prediction for each TEST card. Only with a parent’s or guardian’s permission, you may instead open Teachable Machine in a browser (the link is beside this activity): it trains on your own machine and needs no account, and you never sign in. Make two classes, Round and Sickle, add about 30 samples per class from the TRAIN cards only, press Train, then show each TEST card and record the prediction and confidence.
  7. On your sketchbook page, draw a number line of ratios with the two classes and your threshold, and mark the cards each model got wrong.

Data

CardLength (mm)Width (mm)RatioTrue labelRule predictionMachine prediction
T1
T2 … T10

Analysis

  1. How many TEST cards did your rule get right, and how many did the machine get right? Which cards fooled each, and why?
  2. Did the machine appear to learn shape, or something else such as color or card position? What evidence do you have?
  3. What would you change about the TRAIN cards to make both models more reliable? Connect your answer to the idea that training data carries the habits of whoever made it.
  4. Which moment of this process would be most useful to show a ninth grader who has never heard of machine learning? Sketch it.

If you cannot do this studio in person: Choose one. All three earn full credit, all three take about the same time, and all three end in the same data table, the same number line and the same four analysis questions. Option A, your own cards and a hand-built second model: draw and measure your own 40 cards, then build the two-feature boundary model described below (on paper or on a spreadsheet scatter chart) and test both models on the TEST cards. Option B, your own cards and a browser model, only with a parent’s or guardian’s permission: draw and measure your own 40 cards, then train Teachable Machine (no account, no sign-in) on the TRAIN cards with a webcam or uploaded phone photographs and test it on the TEST cards. Option C, no cards and no computer: use the measured card set printed in the reading for this lesson, compute every ratio, choose your threshold on the 30 TRAIN rows and test it on the 10 TEST rows before reading their labels. Then build a second, different model by hand to compare with your first, as the machine half of the lab: add a second feature, area (length times width), plot the 30 TRAIN cards on a simple two-axis grid, and draw a straight boundary line that separates the labels. That line is a two-feature model. Apply it to the 10 TEST cards and compare its errors with your one-feature rule’s errors. Two of the forty cards defeat any single threshold, and finding them is the heart of this lab.

Formative check

Work through these before moving on. They are not graded — they tell you whether the standard below has landed yet.

Make a prediction

Every sickle card in a training set was drawn in red, and every round card in black. A new round cell is drawn in red. What will a machine-trained model most likely predict?

Quite possibly sickle-shaped. A model learns whatever separates the training labels most easily, and here color does that perfectly. This is called a shortcut or spurious feature.
Fill in the blank

A model that outputs a category is a ; a model that outputs a number performs .

+50 XP

A drawn cell is 15 mm long and 5 mm wide. With the rule ‘if ratio > 2.0, label sickle’, what does the rule predict?

15 / 5 = 3.0, which is greater than 2.0.

Simulation & tools

Train a classifier on your own cell cards

Optional, and only with a parent’s or guardian’s permission. You run this yourself, and you never sign in or create an account. Teachable Machine trains in your own browser; signing in is needed only to save a project, which this lab never asks you to do. The by-hand route at the end of this box earns the same credit. Create two classes, Round and Sickle. Add about 30 samples per class from the TRAIN cards only, using the webcam or by uploading photographs of the cards as image files, then press Train. Show each TEST card and record the prediction and confidence in the data table. Then retrain after removing all the cards drawn in color from one class, and ask: did the predictions on colored TEST cards change? What does that tell you about what the model learned? With no computer at all, take Option C in the alternative and build the two-feature boundary model by hand on the printed card set; it answers the same question.

Open Teachable Machine ↗  ·  Teachable Machine, Google · free to use; Google terms (models train in the browser)

Quick self-check

How confident are you that you can explain supervised learning, features and testing on unseen data from a model I built by hand?

Not yetVery confident

Practice

Work these on paper or in your notebook, then open Check your answer. Aim for all of Fluency and Application; try at least one Challenge.

Printable version: this unit’s practice workbook (PDF)

Fluency

Build speed and accuracy with the core skill.

  1. Define supervised learning, feature and test set in one sentence each.
    Check your answer
    Answer: Supervised learning is machine learning from examples people have already labeled with the correct answer. A feature is a measurable property of an input, such as length-to-width ratio, that a model uses to decide. A test set is labeled examples held back from training and used only to measure performance on unseen data.
    The three ideas connect: labeled examples train the model on features, and the test set checks it on examples it never saw.
  2. Compute the length-to-width ratio, to one decimal place, for these drawn cells: (a) 24 mm by 6 mm; (b) 8 mm by 6 mm; (c) 10 mm by 4 mm.
    Check your answer
    Answer: (a) 4.0; (b) 1.3; (c) 2.5.
    24 / 6 = 4.0; 8 / 6 = 1.33, which rounds to 1.3; 10 / 4 = 2.5.
  3. Classifier or regression? (a) a model that labels a cell drawing round or sickle-shaped; (b) a model that predicts how many days a hospital stay will last; (c) a model that predicts a patient’s heart rate.
    Check your answer
    Answer: (a) classifier; (b) regression; (c) regression.
    A classifier outputs a category; regression outputs a number.

Application

Use the skill in context. Show your reasoning.

  1. In the lesson’s TRAIN set, the round cards have ratios from 0.9 up to 1.4, and the sickle cards have ratios from 1.8 (card C06, 11 mm by 6 mm) up to 4.0. Choose a threshold that sorts all 30 TRAIN cards correctly, and write the rule as a sentence.
    Check your answer
    Answer: Any threshold from 1.4 up to (but not including) 1.8 works, for example 1.5: ‘If ratio > 1.5, label sickle.’
    The highest round ratio is 1.4 (C03, 11 / 8 = 1.375) and the lowest sickle ratio is 1.8 (C06, 11 / 6 = 1.83). A threshold between them separates every TRAIN card.
  2. Apply the rule ‘If ratio > 1.5, label sickle’ to these TEST cards: C35 (14 mm by 12 mm), C36 (17 mm by 5 mm), C38 (17 mm by 8 mm). Then check: the true labels are C35 round, C36 sickle, C38 round. Which card does the rule get wrong?
    Check your answer
    Answer: C35: ratio 1.2, predicted round (correct). C36: ratio 3.4, predicted sickle (correct). C38: ratio 2.1, predicted sickle, but it is round, so C38 is the error.
    14 / 12 = 1.17, about 1.2; 17 / 5 = 3.4; 17 / 8 = 2.125, about 2.1. C38 is a long round cell, one of the deliberately awkward cards.
  3. With the rule ‘If ratio > 2.0, label sickle’, which TRAIN card is misclassified, and which TEST card is misclassified? (C06 is 11 mm by 6 mm, sickle; C38 is 17 mm by 8 mm, round.)
    Check your answer
    Answer: TRAIN: C06 (ratio 1.8, predicted round but sickle). TEST: C38 (ratio 2.1, predicted sickle but round).
    11 / 6 = 1.8, not greater than 2.0, so predicted round. 17 / 8 = 2.1, greater than 2.0, so predicted sickle. These are the two awkward cards: no single threshold gets both right, because C06 needs a threshold below 1.8 and C38 needs one at 2.1 or above.
  4. For the second, two-feature model, the lesson adds area (length times width). Compute the area for C06 (11 by 6) and C38 (17 by 8).
    Check your answer
    Answer: C06: 66 square mm. C38: 136 square mm.
    11 x 6 = 66; 17 x 8 = 136. Plotting ratio against area gives the second axis on which a straight boundary line can be drawn.
  5. Every sickle card in a training set was drawn in red marker. Explain in 2-3 sentences why a machine-trained model might fail on a red round cell, and what you would change about the TRAIN cards.
    Check your answer
    Answer: Model: The model learns whatever separates the training labels most easily, and here color does that perfectly, so it may have learned ‘red’ instead of ‘crescent’. A red round cell would then be labeled sickle. Draw both shapes in both colors, with varied size and pen pressure, so color no longer predicts the label.
    A strong answer names the shortcut (color), explains why the model learned it, and fixes the data rather than the model. Training data carries the habits of whoever made it.

Challenge

Stretch problems. Expect to think before you write.

  1. A student sets the threshold at exactly 1.85 after studying all 40 cards, including the TEST cards, and reports a perfect TRAIN score. Explain the two problems with this and what the lesson calls a rule that fits training data perfectly but fails on new data.
    Check your answer
    Answer: First, looking at the TEST cards before testing means the test no longer measures performance on unseen data. Second, a threshold tuned that tightly may fit quirks rather than the pattern. A rule that fits training data perfectly but fails on new data is overfitted.
    Developers hold back a test set precisely so it stays unseen. A strong answer names both the leak and overfitting.

Review

Keep earlier skills sharp.

  1. Review 1.1: Is the cell-sorting rule you built narrow AI, AGI or superintelligence? Explain in one sentence.
    Check your answer
    Answer: Narrow AI (a tiny narrow model): it performs one specific task, sorting cell drawings into two labels.
    Narrow AI performs a specific task; AGI and superintelligence are hypothetical.

Missouri Learning Standards — Media Arts addressed: MA:Cr1A.I, MA:Cr2A.I, MA:Pr5A.I, RST.11-12.3, RST.11-12.9

Instructional focus: Generative methods and hands-on modeling of a technical process

Reading

Forty Measured Cell Cards: The Data for the Hand-Built Classifier Lab

SLS original teaching data set and worked solution; no external source. Teachable Machine (Google) is named as an optional tool only; retrieved Sept. 26, 2026 · SLS original

Why this reading. Option C of the lab uses “the measured card set printed in this lesson,” and the lesson never printed one. Here it is: forty invented cards, measured, so you can do the whole lab with a pencil, and so the same forty cards can be typed into or photographed for a machine model. SLS original data. These numbers are invented for teaching; they are drawings of cells, not measurements of any patient or any real blood sample.

The card set

Each card is a drawing of a red blood cell, measured in millimeters. Round cards have a pale center; sickle cards are crescent-shaped. The ratio is length ÷ width to one decimal. The area is length × width (a bounding-box area, not the exact area of the shape). The ink color is part of the experiment: in the TRAIN set every sickle card was drawn in red ink and every round card in black, which is exactly the kind of accident the lesson’s prediction question is about.

TRAIN cards (30), labels known

CardLength (mm)Width (mm)Ratio (length ÷ width)Area (mm²)Ink colorTrue label
T01881.064blackround
T02981.172blackround
T03771.049blackround
T041091.190blackround
T05871.156blackround
T06991.081blackround
T071181.488blackround
T08761.242blackround
T0910101.0100blackround
T10861.348blackround
T11971.363blackround
T121782.1136blackround
T131644.064redsickle
T141553.075redsickle
T151443.556redsickle
T161744.268redsickle
T171352.665redsickle
T181863.0108redsickle
T191543.860redsickle
T201243.048redsickle
T211653.280redsickle
T221452.870redsickle
T231343.252redsickle
T241753.485redsickle
T251042.540redsickle
T261562.590redsickle
T271142.844redsickle
T281252.460redsickle
T29881.064blackround
T30991.081blackround

TEST cards (10), labels hidden

Apply your rule to these before you look at the answer key at the end of this reading. Ink color is shown because the machine might use it; a good hand rule should not.

CardLength (mm)Width (mm)RatioArea (mm²)Ink color
E31871.156black
E321091.190red
E33961.554black
E34771.049red
E351543.860black
E361653.280red
E371343.252black
E381362.278red
E39851.640black
E401352.665red

Your one-feature rule

  1. Sort the TRAIN cards by ratio. All the round cards except one sit between 1.0 and 1.4. Look at the round card with the largest ratio and the sickle card with the smallest ratio.
  2. Pick a threshold between them, and write it as a sentence: If ratio > __, label sickle. Count how many of the 30 TRAIN cards it gets right. (Any threshold from 2.2 to 2.3 gets all 30. The gap between the largest round ratio and the smallest sickle ratio is small, so the choice inside it is close to arbitrary, and that is worth remarking on.)
  3. Apply the rule to the ten TEST cards and record each prediction before you open the key.

Your two-feature model

Draw a grid with area (0 to 150 mm²) on the horizontal axis and ratio (0 to 4.5) on the vertical axis. Plot the 30 TRAIN cards, one symbol for round and another for sickle. The cards that no single ratio threshold can separate will look like strays in the plot. Draw one straight line that keeps all round cards on one side and all sickle cards on the other. Then read off the equation of your line, apply it to the TEST cards, and compare the number of errors with your one-feature rule. Any line that separates the 30 TRAIN cards is acceptable; the key gives one that also separates all 40.

If you use a machine model

Teachable Machine is a free browser tool from Google (the address is teachablemachine.withgoogle.com; the lab links it beside the activity). It is a tool you operate, not text you read, so nothing here depends on it. The steps are: (1) start a new image project; (2) name two classes, Round and Sickle; (3) for each class add samples by showing the TRAIN cards to a camera or by uploading photographs of them, about 30 samples per class; (4) press Train and wait for it to finish; (5) show each TEST card, one at a time, and write down the class it names and the confidence percentage it shows. The lab states that this trains in your own browser and does not need an account; if a screen ever asks you to sign in, you can decline, because the lab never asks you to save anything. If the site is unavailable or there is no camera, use the two-feature model above instead. It answers the same question: what did the model learn?

Try the color shortcut. Photograph or draw the TRAIN cards as the table says (sickle red, round black). Then make three extra TEST cards: a round cell drawn in red, a sickle cell drawn in black, and a sickle cell drawn in red but small (8 × 5 mm). If a model calls the red round card “sickle,” it has learned ink color, not shape. The ten TEST cards above already include two red round cards (E32, E34) and three black sickle cards (E35, E37, E39).

Answer key (read after you have made your predictions)

True labels for the TEST cards, and the calls made by three models. The one-feature rule is “ratio > 2.3”; the two-feature model is score = ratio − 0.01 × area, calling a card sickle when the score is above 1.1 (as a line on your grid, ratio = 1.1 + 0.01 × area); the color shortcut calls every red card sickle.

CardTrue labelRatioOne-feature rule (ratio > 2.3)ResultTwo-feature scoreTwo-feature callResultColor shortcut callResult
E31round1.1roundright0.58roundrightroundright
E32round1.1roundright0.21roundrightsicklewrong
E33round1.5roundright0.96roundrightroundright
E34round1.0roundright0.51roundrightsicklewrong
E35sickle3.8sickleright3.15sicklerightroundwrong
E36sickle3.2sickleright2.40sicklerightsickleright
E37sickle3.2sickleright2.73sicklerightroundwrong
E38sickle2.2roundwrong1.39sicklerightsickleright
E39sickle1.6roundwrong1.20sicklerightroundwrong
E40sickle2.6sickleright1.95sicklerightsickleright

What this shows. The one-feature rule gets 8 of 10 right; it misses E38 (a sickle drawn short and fat, ratio 2.2) and E39 (a small sickle, ratio 1.6). Two cards in the whole set, T12 (a large oval round cell, ratio 2.1) and E39 (ratio 1.6), overlap on ratio: no single threshold can be right for both. The two-feature model gets 10 of 10 right and also gets every one of the 30 TRAIN cards right, because it uses the fact that T12 is large and E39 is small. The color shortcut gets 5 of 10, which is what guessing would do. It fits the TRAIN set perfectly and tells you nothing about cells. That is the lesson’s point: a model can score perfectly on its training data and still have learned the wrong thing, and only cards it has not seen can reveal that.

Sources: Teachable Machine (optional tool, Google) (teachablemachine.withgoogle.com)

Up NextComplete & Continue →

Leave a Reply Cancel reply

Explaining the Invisible: Designers Who Made Data Visible
Prev
Critiquing AI Explainers: Accuracy, Clarity and Hype
Next

© 2026 The VR School | Career content by Wiley | Powered by Sejal Learning Systems.

Join free · Accreditation · Privacy Policy · Terms of Service

Modal title

Main Content