Course contentsShow
← Machine Learning and Deep Learning
Lesson 520 of 3,53812. Model Validation and TuningFree lesson

Plotting and Interpreting Learning Curves

Creating learning curves in practice and reading patterns of underfitting, overfitting, and good fit.

Plotting and Interpreting Learning Curves

You'll learn how to create learning curves and identify underfitting, overfitting, and good model fit by reading visual patterns.

What Are Learning Curves in Practice?

Learning curves plot your model's performance (like accuracy or error) on both training and validation sets as training progresses—typically as the number of training examples increases or training iterations accumulate. They're your visual diagnostic tool for understanding how well your model generalizes.

Creating Learning Curves

You plot two lines:

  • Training score: Performance on data the model has seen
  • Validation score: Performance on held-out data

As you add more training examples (or train longer), watch how both scores evolve. This reveals whether your model needs more data, is too complex, or is just right.

Reading the Patterns

Underfitting (High Bias)

Both training and validation scores are low and converge closely. Your model is too simple—like trying to fit a curve with a straight line. It can't capture the pattern even in training data.

Overfitting (High Variance)

Large gap between high training score and low validation score. Your model memorized training data but fails on new examples—like a student who memorized answers without understanding concepts.

Good Fit (Goldilocks Zone)

Both curves converge to a high score with minimal gap. Your model learned the underlying pattern and generalizes well. This is your target!

More Data Needed

If validation score is still improving as you add examples (hasn't plateaued), collecting more data will likely help.

Key Takeaway: Learning curves visually diagnose model health—a wide gap signals overfitting, low convergent scores signal underfitting, and tight convergence at high performance signals a well-tuned model ready for deployment.