📈 Model Quality

Overfitting Lab

Watch a model go from underfitting to just right to badly overfitting — and discover why a model that's too clever on training data fails completely on new data!

📍 Add Data Points

📈 Tune Complexity

📉 Learning Curves

🛡️ Regularisation

🏆 Badge

The Bias-Variance Trade-off

🎯

Underfitting

Model is too simple — misses the real pattern. High training error AND high test error. "The model didn't learn."

✅

Just Right

Model captures the true pattern without memorising noise. Low training error AND low test error. The goal!

🌀

Overfitting

Model memorises training data — even the noise. Very low training error but HIGH test error. "It cheated."

🛡️

Regularisation

Penalise model complexity. Forces simpler solutions. L1, L2, dropout — all ways to prevent overfitting.

📈

Wizzy the AI Tutor

Welcome to the Overfitting Lab! 📈 Click on the canvas to add data points. The points follow a hidden curve with some random noise added. The model's job is to find the curve — but watch what happens when it tries too hard!

Step 1 — Create Your Dataset

0 points

💡 Click to add points · Or use the presets below

🎛️ Controls

Noise level 20

Model degree 2

📐 Degree 2 — Just Right

Error Scores

Train error

—

Test error

—

Add data points and tune the degree slider to see overfitting!

📈

Wizzy the AI Tutor

Now explore model complexity systematically! Click each degree to see what the fitted curve looks like. At degree 1 (a straight line), the model is underfit. At degree 12, it wiggles through every point — classic overfitting! Which degree fits best?

Step 2 — Explore Model Complexity

Click a polynomial degree:

Select a degree

Error Scores

Train error

—

Test error

—

Click a degree to see how complexity affects fit.

📈

Wizzy the AI Tutor

The learning curve is one of the most important tools in ML! It shows training error and validation error as you increase model complexity. The sweet spot is where validation error is at its lowest — before it starts rising again. Drag the slider to explore!

Step 3 — Learning Curves

🟢 Training error (always goes down) 🔴 Validation error (goes up when overfitting!) 🟡 Sweet spot (best degree)

📈

Wizzy the AI Tutor

Regularisation is the fix for overfitting! Instead of making the model simpler, we penalise large coefficients. A regularised degree-12 model produces a much smoother curve — it can't wiggle wildly because that would make the penalty too high. Compare the two sides!

Step 4 — Regularisation: The Fix

Regularisation strength (λ): λ = 0.10

❌ Overfit (Degree 12, no reg.)

Test error: —

✅ Regularised (Degree 12 + λ)

Test error: —

Drag the λ slider to see how regularisation tames the overfitting curve!

📈

Wizzy the AI Tutor

🎊 You've mastered the most important concept in all of ML! Overfitting is why AI models fail in the real world. Every ML engineer fights it daily using train/test splits, learning curves, and regularisation. You now know exactly what they know!

📈

Model Quality Badge!

You mastered overfitting, learning curves, and regularisation!

Optional. Stays on this device only — not sent to WhizzStep.

📈 WhizzStep AI Lab

Activity completion card for

Student Name

has explored Overfitting, Bias-Variance & Regularisation

This records completion of a browser activity only. It is not an accredited certificate or proof of mastery.

Overfitting Expert

Bias-Variance

ML Engineer

whizzstep.in

Key Concepts Mastered

Overfitting

🌀 Memorising Noise

When a model learns the training data too well — including its noise — it fails to generalise to new data.

Underfitting

📏 Too Simple

A model that hasn't learned the pattern at all. High bias. A straight line through clearly curved data.

Train/Test Split

✂️ Held-Out Data

Keep some data hidden from training. Test error on this holdout reveals overfitting before deployment.

Regularisation

🛡️ Complexity Penalty

Add a penalty term to the loss function that grows with coefficient size. Forces simpler, smoother models.

Bias-Variance

⚖️ The Trade-off

High bias = underfit. High variance = overfit. The sweet spot balances both to minimise total error.

Cross-Validation

🔄 K-Fold

Split data into K folds. Train on K-1, test on 1. Repeat K times. More reliable than a single train/test split.

About this lab

AI-10 How AI and Machine Learning Work Classes 8-10 15 min Applied

Learning objective: Fit a model to noisy data and see how it can memorise training examples instead of learning a generalisable pattern.

What this simplifies: A single small synthetic dataset is used; overfitting looks different across real-world datasets and model types.

Privacy: No learner input leaves the device.

Teacher prompt: Ask the class why this simulation might mislead someone who takes it too literally.

Reflect: What is one thing this activity showed you that you did not expect?

← Train My Robot Back to All Labs Cluster Explorer →

Bring this learning to my school →