﻿TEACHER NOTES — PHASE 3: EXCEL BRAIN (60 min)

WHAT IT TEACHES
A neural-network-like learner built only from Excel formulas: pixels x weights + bias -> score -> probability; after each picture the weights are nudged by learning rate x error x pixel (logistic regression trained one picture at a time). Students see learning curves, the learning rate, a "brain scan" of the weights, train/test, and why more varied data helps.

THE WORKBOOK
  Pictures: 16 slots (12 filled). Learn: 192 learning steps (one row each). Brain: dashboard with charts and the 5x5 weight map. Test: 10 pictures never used for training. Draw: free drawing pad. Data: helper (formulas copy pictures into rows).
  Students TYPE: F8, G8, K8, L8, AL8 (then fill right / down), the Test formulas and the summary cells. Everything else is ready.
  Example_ files are the finished versions (use to check, or to show on the projector).

VERIFIED NUMBERS (checked in real Excel and in an independent Python simulation: identical to 1e-6)
  Learning rate 0.5, 12 training pictures: mistakes per round 6, 2, 0, 0 …; loss 0.91, 0.26, 0.13 …; Test accuracy 7/10 = 70% (test pictures 3, 6 and 10 fail).
  Learning rate 0.05: mistakes 5, 4, 3, 2, 1, 0 (slower). Learning rate 5: mistakes reach 0 by round 4, but the loss in round 1 is huge (the brain jumps around), and with the hint pictures it ends at 80%, lower than at 0.5. The biggest learning rate is not automatically the best.
  With the 4 hint pictures added (slots 13-16): 10/10 = 100% at learning rates 0.5 and 1; 9/10 = 90% at 0.05, 0.1 and 2; 8/10 = 80% at learning rate 5.
  The goal is 90%: students can reach it by drawing varied pictures, or by tuning the learning rate.

LESSON FLOW
  0-5 idea (pixels, weights, score). 5-10 workbook tour. 10-25 type the learning formulas, fill down. 25-35 watch the charts, learning-rate experiment, weight map. 35-45 Test sheet: formulas, accuracy 70%. 45-50 Draw sheet: the brain always answers. 50-60 add pictures -> goal; discussion.

KEY POINTS / DISCUSSION
  - Why is the Test sheet separate? (honest evaluation). If students copy test pictures into training the test is meaningless — a good moment to talk about leakage.
  - The brain only knows pixel POSITIONS: a banana drawn upright, or an apple moved by one square, confuses it. Real image AIs (like Teachable Machine) look for shapes anywhere.
  - Learning rate: too small = slow, too big = jumpy/overconfident.
  - Connect to Phase 1: the Fix Pack and the hint pictures do the same thing: more variety.

TROUBLESHOOTING
  Brain sheet says "fill the Learn sheet first": Learn formulas missing, or not filled down to row 199.
  #NAME?/#REF!: the sheet name in a formula is wrong (click cells instead of typing sheet names). The Arabic and Hebrew workbooks have localized sheet names.
  Excel for the web works. Google Sheets works for the formulas but conditional-format colours and charts differ.
