Learning Curve
Last Updated: July 29, 2026 | By Mihail Sebastian | AI Dictionary
A plot of model error against training progress or dataset size, read to diagnose whether a model is overfitting, underfitting, or still improving.
What is a Learning Curve?
A learning curve is a plot of a model’s error against training progress, drawn as two lines: one for the training data and one for a held-out validation set. The distance and direction of those two lines diagnose what the model is doing.
The x-axis is either training iterations (epochs) for one training run, or training-set size across runs. The second variant answers a budget question: would collecting more data help?
How to Read a Learning Curve
- Both lines high and close together: the model fails even on its own training data, which is underfitting. More data will not help; more capacity or better features will.
- Training line low, validation line high and diverging: the model fits its training data but not the world, which is overfitting. Regularization, more data, or stopping sooner closes the gap.
- Both lines descending and converging: healthy learning. If the validation line is still falling at the end of the plot, the model has not finished improving.
The curve turns model evaluation from a single score into a trajectory. Two models with identical final accuracy look very different when one is converged and the other is still climbing.
Example of a Learning Curve
A team trains a sentiment classifier for 50 epochs and plots both losses. Training loss falls the entire run. Validation loss falls with it until epoch 20, flattens, then climbs from epoch 25 onward.
Reading the plot: everything the model learned after epoch 20 hurt generalization, so the run overfit for its last 30 epochs. The fix is written on the chart – retrain with early stopping set to end near epoch 20, or add regularization and check whether the divergence disappears. Without the curve, the team would only have seen a mediocre final score, with no indication of when things went wrong.
Related AI terms: Overfitting · Underfitting · Early Stopping · Cross-Validation · Epoch
Did you like the Learning Curve gist?
Learn about 250+ need-to-know artificial intelligence terms in the AI Dictionary.
Mihail Sebastian — Writes about AI governance, regulation, and the technology behind them. Placeholder bio — replace with a real credential line. About