🧠 Day 12

Memorizer vs Learner

Yesterday your AI aced its practice but scored lower on the exam. Why? Because there are two ways to "pass": memorize every answer (works only on stuff you've seen) or actually learn the pattern (works on anything). Let's meet both.

↓ let's compare
🏷️ The pro wordMemorizing the training data is overfitting; being too simple to learn it is underfitting. The aim is generalization, and the practice-vs-exam gap is the generalization gap (the bias-variance tradeoff).
🧠 Lab · Pick a Brain

Too simple, just right, or too wiggly?

The dots are two teams (🔵 and 🟣) that make an X-shape — you can't split them with one straight line! Try three "brains" and watch two scores: Practice (data it learned) and Exam (fresh data). The gap between them tells the whole story.

📘 PRACTICE
📝 EXAM
💡 The Two Traps

Underfitting vs Overfitting

📏

Underfitting

Brain too simple — it can't even learn the practice data. Both scores low. Like studying too little.

Just Right

Learns the real pattern. Practice and exam both good, small gap. This is what you want!

🌀

Overfitting

Memorizes every dot, even the mistakes. Practice ~100%, exam way lower. A big gap is the warning sign.

🔎 The tell-tale sign: a big gap between practice and exam scores means your AI is memorizing, not learning. Fix it with more data, a simpler brain, or augmentation (Day 13!).
🚀 Beyond the Basics

How the pros really do it

Picking 'just right' complexity is the idea. Pros have real tools to prevent overfitting:

🟢

Tune complexity (what you did)

✅ Shows the tradeoff. ⚠️ By hand only.

🛡️

Regularization

Dropout, weight decay, early stopping - nudge the model to stay simple and general.

📊

More data + validation

More varied data and a validation set catch overfitting before it ships.

🏷️ Pro names to look upoverfitting/underfitting · bias-variance · dropout / weight decay / early stopping.

🎉 Day 12 Complete!

You can now spot a memorizer: watch the gap between practice and exam scores. A real learner generalizes to new stuff.

Day 13: Make Your Data Tougher →
Tomorrow: one of the best ways to stop memorizing and start learning — data augmentation: making your practice data as messy as the real world. 🌪️
AI Adventure · Module 1 · Day 12 — made for young inventors 🚀