What Learning Curves Reveal
You'll learn: How plotting training and validation performance over time helps diagnose whether your model is underfitting, overfitting, or just right.
The Core Idea
A learning curve is a graph that shows how your model's performance (like accuracy or loss) changes as training progresses—typically measured across epochs or training set size. You plot two lines: one for training performance and one for validation performance.
Think of it like learning to ride a bike: if you only practice on flat ground (training data) and can't handle hills (validation data), you've memorized the flat path but haven't learned general biking skills. Learning curves reveal this gap.
What the Curves Tell You
1. Overfitting (High Variance)
- Training performance keeps improving
- Validation performance plateaus or worsens
- Large gap between the two curves
- Fix: Add regularization, get more data, reduce model complexity
2. Underfitting (High Bias)
- Both training and validation performance are poor
- Curves converge but at a low performance level
- Fix: Use a more complex model, add features, train longer
3. Good Fit
- Both curves improve and converge at high performance
- Small gap between training and validation
- Goal achieved: your model generalizes well
4. More Data Needed
- Curves are still improving and haven't plateaued
- Training and validation curves are getting closer
- Action: Collect more training samples
Key Takeaway: Learning curves visually diagnose model health—wide gaps signal overfitting, low convergence signals underfitting, and converging curves at high performance signal a well-tuned model ready for deployment.