Englishالعربيةעברית

Phase 1 — Boss Pictures

Much harder pictures. Train an AI, find out where it fails, then fix it.

About 60 minutes · Teachable Machine (no camera) · Student pack + Fix Pack

🎬 Watch the video first

The video does not travel inside this ZIP (it is large). Open this guide from the class server (http://YOUR-IP:8000/boss/guide-en.html) or online at https://schoolmodel.vercel.app/boss/guide-en.html to watch it.

🚀 Your mission: earn 5 stars

Answer the quick checks in the guide. Each correct answer lights a star.

⭐⭐⭐⭐⭐ 0/5

What is new this time?

  1. The pictures come from messy real life: different tables, dim light, tiny fruit, cut-off fruit — even a pencil sketch.
  2. Your AI will look great while it trains. Then it meets the Boss test.
  3. You play detective: why does the AI fail on some pictures?
  4. Then you fix it with a Fix Pack and train again (Round 2). Did it get better?

Before you start

  1. A computer with internet. Open https://teachablemachine.withgoogle.com/ in Chrome.
  2. Unzip the student pack. Do not open the teacher files.
  3. Keep the results sheet next to you (the file RESULTS_SHEET_EN.txt) — you will use your numbers again in Phase 2 (Excel).

0The mission

Last time the test pictures were easy. This time each picture is a small trap. Your AI must learn from 120 training pictures and then face three sets of test pictures it has never seen.

1_TRAIN120 pictures to teach the AI: Apple, Banana, Orange (40 each)
2_BOSS_TEST_1212 Boss pictures. The file name tells you the real fruit (apple_b01 …).
3_NIGHTMARE_1212 very hard pictures. The file name does NOT tell you the answer.
4_TRAPS_1212 pictures that are NOT apple, banana or orange. Nothing tells you the answer — look at them yourself.
Training pictures (1_TRAIN)
Training pictures (1_TRAIN)
Boss test (2_BOSS_TEST_12)
Boss test (2_BOSS_TEST_12)
Nightmare (3_NIGHTMARE_12)
Nightmare (3_NIGHTMARE_12)
Traps (4_TRAPS_12)
Traps (4_TRAPS_12)
Never train on test pictures. If you do, the AI has seen the answers and the test means nothing.

1Open Teachable Machine and make 3 classes

  1. Go to teachablemachine.withgoogle.com → Get Started → Image Project → Standard image model.
  2. Click the pencil next to Class 1 and type Apple. Rename Class 2 to Banana. Click Add a class and name it Orange.
  3. In the Apple class click Upload → Choose images from your files. Open the folder 1_TRAIN/Apple, press Ctrl+A and open. Wait until it says 40 Image Samples.
  4. Do the same for Banana (folder 1_TRAIN/Banana) and Orange (folder 1_TRAIN/Orange).
Teachable Machine with the three classes uploaded.
Teachable Machine with the three classes uploaded.
Do not use the Webcam. Do not upload anything from the other folders yet.
✔ Each class shows 40 Image Samples.

2Train — and make a prediction first

  1. Before you press the button, write your guess in the results sheet: how many of the 12 Boss pictures will the AI get right?
  2. Click Train Model. Keep the tab open until you see Model Trained (1–2 minutes).
After training: Model Trained.
After training: Model Trained.
Do not switch to another tab while it trains. Training runs inside your browser.
✔ The Training box says Model Trained and the Preview box shows three bars: Apple, Banana, Orange.
🎯 Quick check

Why do we NOT train on the Boss test pictures?

3Boss test: 12 pictures

  1. In the Preview box change Input from Webcam to File.
  2. Click Choose images from your files and open the first picture from 2_BOSS_TEST_12.
  3. Look at the three bars. Write the three percentages in the results sheet and write what the AI answered (the biggest bar).
  4. Repeat for all 12 pictures. Count how many the AI got right.
A prediction in the Preview box: the three percentages are what you write down.
A prediction in the Preview box: the three percentages are what you write down.
Write the numbers exactly as Teachable Machine shows them. You will type them into Excel in Phase 2.
✔ You have 12 rows filled: Apple %, Banana %, Orange %, AI answer, correct or not.
🎯 Quick check

The AI shows Orange 95% for a picture. What does 95% mean?

4Human vs AI: the Nightmare pictures

  1. Open the folder 3_NIGHTMARE_12. For each picture write your own answer first: Apple, Banana, Orange or Mixed (more than one fruit).
  2. Then test the picture in Teachable Machine and write the AI's answer and its biggest percentage.
  3. Score: you get 1 point for every picture you got right. The AI gets 1 point for every picture it got right. Who wins?
✔ You know your score and the AI's score.

5The Trap pictures

The pictures in 4_TRAPS_12 are tomatoes, lemons, a peach, a pumpkin, a ball… They are not apple, banana or orange.

  1. Test at least 6 trap pictures. Write the AI's answer and the biggest percentage.
  2. Question: how sure was the AI? Was it ever unsure?
Your AI only knows three answers, so it must pick one. It has no way to say “none of these”. That is a real problem for AI in the real world.
🎯 Quick check

A tomato is shown to an Apple/Banana/Orange AI. What must it do?

6Detective time

Look at the pictures the AI got wrong. What do they have in common? Tick every clue that fits:

Write your 3 best clues in the results sheet. A good detective gives evidence, not guesses.
🎯 Quick check

The AI fails on fruit lying on unusual tables. What is the best fix?

7Round 2: the Fix Pack

  1. Ask your teacher for the Fix Pack (a ZIP file). Unzip it.
  2. In Teachable Machine upload the pictures from FIX/Apple into the Apple class, FIX/Banana into Banana and FIX/Orange into Orange (20 more each).
  3. Click Train Model again. Test the same 12 Boss pictures again and write the new numbers in the Round 2 columns.
  4. Compare Round 1 and Round 2. How many more pictures are right now?
Fix Pack pictures
Fix Pack pictures
✔ You have two scores for the Boss test: Round 1 and Round 2.
🎯 Quick check

After the Fix Pack the score went up. Which pictures helped?

8Extension: add a fourth class called Other

Can we teach the AI to say “this is none of the three”?

  1. Click Add a class and name it Other.
  2. Upload the 24 pictures from the folder OTHER_PACK/Other (tomatoes, lemons, balls …).
  3. Train again and test the 12 trap pictures. How many are called Other now?
OTHER_PACK pictures
OTHER_PACK pictures
Teaching an AI what it should NOT answer is a trick real engineers use. Notice the traps in the pack are different pictures from the traps you test with.

9Think about it

  1. Why did the AI look perfect during training but fail on some Boss pictures?
  2. Which kind of picture was the hardest for your AI? Why do you think so?
  3. Who won the Nightmare round: you or the AI? What does that tell you?
  4. What did the Fix Pack change? Is it enough? What else would you add?
  5. Would you trust this AI to sort fruit in a shop? What could go wrong?

Next stop: take your Round 1 and Round 2 numbers to Excel in Phase 2.

⭐ Boss challenge

  1. Reach 10 of 12 on the Boss test in Round 2.
  2. Get more points than the AI in the Nightmare round.
  3. Find one picture where the AI is more than 90% sure and wrong. Show it to the class.

Words to know

Englishالعربيةעברית
training picturesصور التدريبתמונות אימון
test picturesصور الاختبارתמונות מבחן
classفئةמחלקה
accuracyالدقةדיוק
confidenceالثقةביטחון
shortcutاختصار (طريقة غير مقصودة للتعلّم)קיצור דרך
biasانحيازהטיה
Fix Packحزمة الإصلاحחבילת תיקון
trap pictureصورة فخתמונת מלכודת

Questions

It says there are too few samples.

Teachable Machine needs at least a few pictures in every class. Check that each class says 40 Image Samples.

My numbers are different from my friend's.

That is normal. Training has some randomness, so two students can get slightly different results. Compare the patterns, not exact numbers.

Do I need a Google account?

No. Do not click Save to Drive. Teachable Machine works inside your browser.

The AI got everything right in Round 1.

Check that you used the Boss test folder, not the training folder. If so, tell your teacher — and try the Nightmare and Trap pictures.